CoChat
Sign Up
Cited by 58Advanced Neural Network ApplicationsAutonomous Vehicle Technology and SafetyDomain Adaptation and Few-Shot Learning

Multimodal semantic segmentation in autonomous driving: A review of current approaches and future perspectives

Open access
Rizzoli, Giulia, Barbato, Francesco, Zanuttigh, Pietro·2022·Technologies
DOI 10.3390/technologies10040090
Open this paper in CoChat — free accountOpen PDF

Abstract

The perception of the surrounding environment is a key requirement for autonomous driving systems, yet the computation of an accurate semantic representation of the scene starting from RGB information alone is very challenging. In particular, the lack of geometric information and the strong dependence on weather and illumination conditions introduce critical challenges for approaches tackling this task. For this reason, most autonomous cars exploit a variety of sensors, including color, depth or thermal cameras, LiDARs, and RADARs. How to efficiently combine all these sources of information to compute an accurate semantic description of the scene is still an unsolved task, leading to an active research field. In this survey, we start by presenting the most commonly employed acquisition setups and datasets. Then we review several different deep learning architectures for multimodal semantic segmentation. We will discuss the various techniques to combine color, depth, LiDAR, and other modalities of data at different stages of the learning architectures, and we will show how smart fusion strategies allow us to improve performances with respect to the exploitation of a single source of information.

AI snapshot

Written from the paper’s full text.

TL;DR

This survey reviews deep learning approaches for multimodal semantic segmentation in autonomous driving, covering sensors, datasets, and fusion architectures. It finds that combining modalities generally improves segmentation over single-source input, but labeled multimodal data, especially thermal and raw LiDAR fusion, remain limiting.

Key findings

  • Multimodal fusion improves performance over using a single modality; common designs use one network branch per modality plus modules that merge or exchange features across branches.
  • Fusion strategies are categorized as early, feature-level, and late fusion, with recent methods often performing multiple fusion operations at different network stages.
  • Reported benchmark mIoU values include LWM at 83.4 and SSMA at 83.29 on Cityscapes (2.5D SS), PMF at 63.9 on KITTI (2D + 3D SS), and CMX at 59.7 on MSSSD/MF (RGB + Thermal SS).
  • Data availability is a major bottleneck: few large-scale multimodal semantic segmentation datasets exist, and no standard large-scale thermal dataset is available.

Methods

  • Design: narrative survey/review of deep learning architectures for multimodal semantic segmentation in autonomous driving.
  • Scope: acquisition setups and sensors (RGB, thermal, depth, stereo, ToF, LiDAR, RADAR, GPS/IMU), datasets, and fusion techniques for RGB+depth, RGB+LiDAR, pointcloud segmentation, and other modalities.
  • Analysis: compares methods by modalities, datasets, fusion approach, fusion location, and reports mIoU on Cityscapes, KITTI, and MSSSD/MF; no new experiments are conducted.

Limitations

  • Stated: labeled data availability is a bottleneck, especially for semantic segmentation; thermal data lack a standard large-scale dataset; RGB+LiDAR fusion using raw pointclouds and RADAR fusion remain largely unexplored.
  • Stated: there is no indication that one fusion scheme is preferable to others, and current metrics compare semantic accuracy rather than multimodal resilience.
  • Evident from the text: as a narrative survey, it does not describe a systematic search or selection protocol and does not perform a quantitative meta-analysis.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/technologies10040090

Cites

129 of 129 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 104 cited sources checked · 0 supported · 0 flagged

  • Premisereasoning

    Computing an accurate semantic representation of a scene from RGB information alone is very challenging.

    “the computation of an accurate semantic representation of the scene starting from RGB information alone is very challenging.”
    Quote in the paperSubstantial claim
  • Premisereasoning

    Depth measurements provide useful geometric cues that help significantly in the discrimination of visual ambiguities.

    “Nevertheless, depth measurements provide useful geometric cues, which help significantly in the discrimination of visual ambiguities, e.g., to distinguish between two objects with a similar appearance.”
    Quote in the paperSubstantial claim
  • Interpretationreasoning

    Early fusion of RGB and other modalities into multi-channel representations fed into classical encoder-decoder networks is not too effective because it struggles to capture the different type of information carried by different modalities.

    “This simple early fusion combination strategy is not too effective because it struggles to capture the different type of information carried by the different modalities (e.g., RGB images contain color and texture, whereas the other modalities typically better represent the spatial relations among objects).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [128] Couprie, C., Farabet, C., Najman, L., and LeCun, Y. (2013). Indoor semantic segmentation using depth information. arXiv.

      Not checked in the free run.

    • cites [129] Pagnutti, 2017, Segmentation and semantic labelling of RGBD data with convolutional neural networks and surface fitting, IET Comput. Vis. source

      Not checked in the free run.

  • Premisecitation

    SSMA, using a ResNet50 backbone, achieves an mIoU of 83.29 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| SSMA [98] | ResNet50 [16] | 83.29 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [98] Valada, 2020, Self-supervised model adaptation for multimodal semantic segmentation, Int. J. Comput. Vis. source

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.
  • Premisecitation

    CMX, using a MiT-B4 backbone, achieves an mIoU of 82.6 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| CMX [99] | MiT-B4 [27] | 82.6 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [99] Liu, H., Zhang, J., Yang, K., Hu, X., and Stiefelhagen, R. (2022). CMX: Cross-Modal Fusion for RGB-X Semantic Segmentation with Transformers. arXiv. source

      Not checked in the free run.

    • cites [27] Xie, 2021, SegFormer: Simple and efficient design for semantic segmentation with transformers, Adv. Neural Inf. Process. Syst.

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.
  • Classified as probably not a substantial claim. 48
  • May combine two claims. 22
  • The quote only partly states this claim. 1

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Vision Transformers present a convolution-free, transformer-based vision approach able to surpass previous state-of-the-art techniques in image classification at the cost of much higher memory and training data requirements.

    “This architecture has been adapted to the image understanding field in the Vision Tranformers (ViT) [24,25] work, which presents a convolution-free, transformer-based vision approach able to surpass previous state-of-theart techniques in image classification (at the cost of much higher memory and training data requirements).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [24] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., and Gelly, S. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv.

      Not checked in the free run.

    • cites [25] Kolesnikov, A., Dosovitskiy, A., Weissenborn, D., Heigold, G., Uszkoreit, J., Beyer, L., Minderer, M., Dehghani, M., Houlsby, N., and Gelly, S. (2021, January 3–7). An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations (ICLR), Virtual Event, Austria.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DeepLab adopts pyramid pooling modules wherein feature maps are implicitly downsampled through dilated convolutions of different rates.

    “Inspired by the FPN model, Chen et al. [21,22] proposes the DeepLab architecture, which adopts pyramid pooling modules wherein the feature maps are implicitly downsampled through the use of dilated convolutions of different rates.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [21] Chen, 2017, Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, IEEE Trans. Pattern Anal. Mach. Intell. source

      Not checked in the free run.

    • cites [22] Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., and Adam, H. (2018, January 8–14). Encoder-decoder with atrous separable convolution for semantic image segmentation. Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    PointNet was one of the first general-purpose 3D pointcloud segmentation architectures, but its performance decreases significantly in outdoor settings due to sparse LiDAR data, limiting its applicability in autonomous driving.

    “More in detail, PointNet [29,30] was one of the first general-purpose 3D pointcloud segmentation architectures, but although it achieved state-of-the-art results on indoor scenes, the sparse nature of LiDAR data led to a significant performance decrease in outdoor settings, limiting its applicability in autonomous driving scenarios.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [29] Qi, C.R., Su, H., Mo, K., and Guibas, L.J. (2017, January 21–26). Pointnet: Deep learning on point sets for 3d classification and segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA.

      Not checked in the free run.

    • cites [30] Qi, C.R., Yi, L., Su, H., and Guibas, L.J. (2017, January 4–9). Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Proceedings of the 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    LiDARs provide very high accuracy but have major downsides: high cost, delicate moving parts, and sparse depth maps rather than dense images.

    “Although such sensors provide very high accuracy, they come with a couple of major downsides, namely the high cost, the presence of delicate moving parts, and the fact that the depth map produced is sparse, rather than dense as the images from standard cameras.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Dilated convolutions allow for an exponential increase in the receptive field without a decrease in resolution or increase in parameters.

    “According to the authors, dilated convolutions allow for an exponential increase in the receptive field without a decrease in resolution or increase in parameters, as may happen in the traditional pooling or stride-based approaches.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [21] Chen, 2017, Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, IEEE Trans. Pattern Anal. Mach. Intell. source

      Not checked in the free run.

    • cites [22] Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., and Adam, H. (2018, January 8–14). Encoder-decoder with atrous separable convolution for semantic image segmentation. Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisereasoning

    Most autonomous cars exploit a variety of sensors, including color, depth or thermal cameras, LiDARs, and RADARs.

    “For this reason, most autonomous cars exploit a variety of sensors, including color, depth or thermal cameras, LiDARs, and RADARs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    FCN introduced an end-to-end convolutional model made of an encoder and a decoder.

    “The first approach to showcase the deep learning potential on this task is found in [14], which introduced an end-to-end convolutional model, the so-called Fully Convolutional Network (FCN) model, which is made of an encoder (or contraction segment) and a decoder (or expansion segment).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [14] Long, J., Shelhamer, E., and Darrell, T. (2015, January 7–12). Fully convolutional networks for semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, USA. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    FPN uses a bottom-up pathway, a top-down pathway, and lateral connections to join low-resolution and high-resolution features and better propagate low-level information.

    “A slightly different approach is proposed in the Feature Pyramid Network (FPN), developed by Lin et al. [20], where a bottom-up pathway, a top-down pathway, and lateral connections are used to join low-resolution and high-resolution features and to better propagate the low-level information into the network.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [20] Lin, T.Y., Dollár, P., Girshick, R., He, K., Hariharan, B., and Belongie, S. (2017, January 21–26). Feature pyramid networks for object detection. Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    RGB cameras are sensitive to light and weather conditions which can lead to failures in outdoor environments.

    “Moreover, RGB cameras are sensitive to light and weather conditions which can lead to failures in outdoor environments [37].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [37] Secci, F., and Ceccarelli, A. (2020, January 12–15). On failures of RGB cameras and their effects in autonomous driving applications. Proceedings of the IEEE 31st International Symposium on Software Reliability Engineering (ISSRE), Coimbra, Portugal. source

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    The lack of geometric information and the strong dependence on weather and illumination conditions introduce critical challenges for approaches tackling semantic segmentation from RGB alone.

    “In particular, the lack of geometric information and the strong dependence on weather and illumination conditions introduce critical challenges for approaches tackling this task.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    DeconvNet proposes overlapping deconvolution and unpooling layers to perform nonlinear upsampling, improving performance at the cost of increasing training complexity.

    “DeconvNet (Noh et al. [19]) which proposes an architecture that contains overlapping deconvolution and unpooling layers to perform nonlinear upsampling, resulting in improving the performance at the cost of increasing the complexity of the training procedure.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Noh, H., Hong, S., and Han, B. (2015, January 11–18). Learning deconvolution network for semantic segmentation. Proceedings of the IEEE International Conference on Computer Vision, Santiago, Chile. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The transformer architecture was introduced by Vaswani et al. in 2017 to reduce the dependence of NLP architectures on recurrent blocks.

    “The most famous approach of this class is the transformer architecture [23], introduced by Vaswani et al. in 2017 in an effort to reduce the dependence of NLP architectures on recurrent blocks, which have difficulty in handling long-time relationships between input data.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., and Polosukhin, I. (2017, January 4–9). Attention is all you need. Proceedings of the 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Smart fusion strategies allow improving performance with respect to the exploitation of a single source of information.

    “we will show how smart fusion strategies allow us to improve performances with respect to the exploitation of a single source of information.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Thermal cameras give temperature-based characteristics of objects, which can better enhance recognition of some objects, thereby improving resilience of semantic scene understanding in challenging lighting conditions.

    “Thermal cameras give temperature-based characteristics of the objects, which can better enhance the recognition of some objects, thereby improving the resilience of semantic scene understanding in challenging lighting conditions [38].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] Gade, 2014, Thermal cameras and applications: A survey, Mach. Vis. Appl. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Early approaches to semantic segmentation were based on classifiers on small image patches until deep learning enabled great improvements.

    “Early approaches to semantic segmentation were based on the use of classifiers on small image patches [11][12][13], until the introduction of deep learning, which has enabled great improvements in this field as well.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [11][12][13]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    How to efficiently combine all sensor sources to compute an accurate semantic description of the scene is still an unsolved task.

    “How to efficiently combine all these sources of information to compute an accurate semantic description of the scene is still an unsolved task, leading to an active research field.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    ParseNet models global context directly rather than only relying on a larger receptive field.

    “Other remarkable architectures that followed FCN are ParseNet (Liu et al. [18]), which models global context directly rather than only relying on a larger receptive field”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18] Liu, W., Rabinovich, A., and Berg, A.C. (2015). Parsenet: Looking wider to see better. arXiv.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    RandLANet adds a grid-based downsampling step and feature aggregation based on random-centered KD-trees to better handle sparse LiDAR samples.

    “An evolution of this technique is developed in RandLANet [31], where an additional grid-based downsampling step is added as preprocessing, together with a feature aggregation based on random-centered KD-trees, to better handle the sparse nature of LiDAR samples.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., and Markham, A. (2020, January 13–19). Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    The joint employment of various data streams from sensors allows a much more in-depth understanding of the environment.

    “The joint employment of the various data streams coming from the sensors (RGB, LiDAR, RADAR, stereo setups, etc.) allows a much more in-depth understanding of the environment.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    DDAD provides semantic segmentation labels only for the validation and test (non-public) sets, significantly restricting its use-case.

    “The dataset provides semantic segmentation labels only for the validation and test (non-public) sets, significantly restricting its use-case.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [82] Guizilini, V., Ambrus, R., Pillai, S., Raventos, A., and Gaidon, A. (2020, January 14–19). 3D Packing for Self-Supervised Monocular Depth Estimation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    WoodScape's 2D information is extracted only by using fish-eye cameras.

    “As opposed to its competitors, its 2D information is extracted only by using fish-eye cameras.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [84] Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Varley, P., O’Dea, D., Uricár, M., Milz, S., Simon, M., and Amende, K. (November, January 27). Woodscape: A multi-task, multi-camera fisheye dataset for autonomous driving. Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Korea.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    RGB+LiDAR fusion is a field where data is abundant but that is still mostly unexplored due to the significant modality difference, especially when exploiting the LiDAR samples as raw pointclouds and not after projection.

    “On the other hand, a field where data is abundant but that is still mostly unexplored (due to the significant modality difference) is RGB+LiDAR fusion, especially when exploiting the LiDAR samples as raw pointclouds and not after projection.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Current metrics compare the networks' accuracy on the semantic prediction directly rather than considering multi-modal resilience.

    “In turn, current metrics compare the networks' accuracy on the semantic prediction directly rather than considering multi-modal resilience.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Synthia is one of the oldest multimodal synthetic datasets providing labeled semantic segmentation samples.

    “Synthia [70-72] is one of the oldest multimodal synthetic datasets providing labeled semantic segmentation samples.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [70] Ros, G., Sellart, L., Materzynska, J., Vazquez, D., and Lopez, A.M. (2016, January 27–30). The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA. source

      Not checked in the free run.

    • cites [71] Hernandez-Juarez, D., Schneider, L., Espinosa, A., Vazquez, D., Lopez, A.M., Franke, U., Pollefeys, M., and Moure, J.C. (2017, January 4–7). Slanted Stixels: Representing San Francisco’s Steepest Streets. Proceedings of the British Machine Vision Conference (BMVC), London, UK. source

      Not checked in the free run.

    • cites [72] Zolfaghari Bengar, J., Gonzalez-Garcia, A., Villalonga, G., Raducanu, B., Aghdam, H.H., Mozerov, M., Lopez, A.M., and van de Weijer, J. (November, January 27). Temporal Coherence for Active Learning in Videos. Proceedings of the IEEE International Conference in Computer Vision, Workshops (ICCV Workshops), Seoul, Korea.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Cityscapes became one of the most common semantic segmentation datasets for autonomous driving benchmarks.

    “Cityscapes [68] became one of the most common semantic segmentation datasets for autonomous driving benchmarks.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [68] Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., and Schiele, B. (2016, January 21–26). The cityscapes dataset for semantic urban scene understanding. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AsymFusion, using an Xception65 backbone, achieves an mIoU of 82.1 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| AsymFusion [100] | Xception65 [122] | 82.1 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [100] Wang, Y., Sun, F., Lu, M., and Yao, A. (2020, January 12–16). Learning deep multimodal feature representation with asymmetric multi-layer fusion. Proceedings of the 28th ACM International Conference on Multimedia, Seattle, WA, USA. source

      Not checked in the free run.

    • cites [122] Chollet, F. (2017, January 21–26). Xception: Deep Learning with Depthwise Separable Convolutions. Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Multiple decoders can also be used but it is a less common choice.

    “Multiple decoders can also be used but it is a less common choice [121].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [121] Jaritz, M., Vu, T.H., Charette, R.d., Wirbel, E., and Pérez, P. (2020, January 13–19). xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Passive stereo camera systems are one of the most common and cost-effective approaches for depth estimation.

    “Passive Stereo camera systems are one of the most common and cost-effective approaches for depth estimation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Even if color and 3D data are the two key sources of information for semantic understanding, other imaging techniques have also been exploited in combination with them.

    “Even if color and 3D data are the two key sources of information for semantic understanding, other imaging techniques have also been exploited in combination with them.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    Semantic KITTI has rapidly become one of the most common benchmarks for LiDAR semantic segmentation.

    “It has rapidly become one of the most common benchmarks for LiDAR semantic segmentation, especially thanks to the significant number of samples made available.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [35] Behley, J., Garbade, M., Milioto, A., Quenzel, J., Behnke, S., Stachniss, C., and Gall, J. (November, January 27). Semantickitti: A dataset for semantic scene understanding of lidar sequences. Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Korea.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    The segmentation map is typically computed by a decoder taking as input the fused features and/or the output of some of the branches.

    “The segmentation map is typically computed by a decoder taking in input the fused features and/or the output of some of the branches.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    The search for an optimal fusion architecture is often driven by empirical results.

    “The search for an optimal fusion architecture is often driven by empirical results.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationnone given

    Hierarchical fusion strategy leverages multilevel features via progressive feature merging and generates a refined feature map, entailing fusing features at various levels rather than at early or late stages.

    “This hierarchical fusion strategy leverages multilevel features via progressive feature merging and generate a refined feature map.It entails fusing features at various levels rather than at early or late stages.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The data bottleneck is particularly critical for semantic segmentation wherein labeling is extremely costly and the task itself is notably data-hungry.

    “This is particularly critical for semantic segmentation wherein labeling is extremely costly and the task itself is notably data-hungry.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    LiDAR is a long-range, omnidirectional depth sensor with high robustness in geometry acquisition at higher cost.

    “A LiDAR is a long-range, omnidirectional depth sensor, which comes with high robustness in geometry acquisition at the expense of a higher cost [39].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [39] Testolina, P., Barbato, F., Michieli, U., Giordani, M., Zanuttigh, P., and Zorzi, M. (2022). SELMA: SEmantic Large-scale Multimodal Acquisitions in Variable Weather, Daytime and Viewpoints. arXiv. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    SA-Gate, using a ResNet101 backbone, achieves an mIoU of 81.7 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| SA-Gate [101] | ResNet101 [16] | 81.7 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [101] Chen, X., Lin, K.Y., Wang, J., Wu, W., Qian, C., Li, H., and Zeng, G. (2020, January 23–28). Bi-directional cross-modality feature propagation with separation-and-aggregation gate for RGB-D semantic segmentation. Proceedings of the European Conference on Computer Vision, Glasgow, UK. source

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • May combine two claims.

    Premisenone given

    Standard color cameras do not provide distance information, making it impossible to access precise information about the positions and sizes of objects.

    “First of all, they do not provide distance information, making it impossible to access precise information about the positions and sizes of the objects.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Virtual KITTI does not provide labels for the LiDAR pointclouds.

    “Unfortunately, the dataset does not provide labels for the LiDAR pointclouds.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [73] Gaidon, A., Wang, Q., Cabon, Y., and Vig, E. (2016, January 21–26). Virtual Worlds as Proxy for Multi-Object Tracking Analysis. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA. source

      Not checked in the free run.

    • cites [74] Cabon, Y., Murray, N., and Humenberger, M. (2020). Virtual kitti 2. arXiv.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    KITTI 360 offers very limited environmental variability.

    “As such, it offers very limited environmental variability.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [83] Liao, Y., Xie, J., and Geiger, A. (2021). KITTI-360: A Novel Dataset and Benchmarks for Urban Scene Understanding in 2D and 3D. arXiv.

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Many real and synthetic datasets have been introduced, but they are still far from being able to represent all the situations that can appear in a real-world driving scenario.

    “Many of them have been introduced, but they are still far from being able to represent all the situations that can appear in a real-world driving scenario.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    SELMA is a synthetic dataset developed in a modified CARLA simulator whose goal is to provide multimodal data in a multitude of environmental conditions, while also allowing a researcher to control such conditions.

    “is a very recent (2022) synthetic dataset developed in a modified CARLA simulator [96] whose goal is to provide multimodal data in a multitude of environmental conditions, while also allowing a researcher to control such conditions.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [39] Testolina, P., Barbato, F., Michieli, U., Giordani, M., Zanuttigh, P., and Zorzi, M. (2022). SELMA: SEmantic Large-scale Multimodal Acquisitions in Variable Weather, Daytime and Viewpoints. arXiv. source

      Not checked in the free run.

    • cites [96] Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., and Koltun, V. (2017, January 13–15). CARLA: An Open Urban Driving Simulator. Proceedings of the 1st Annual Conference on Robot Learning, Mountain View, CA, USA.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    LiDAR pointclouds are relatively sparse and lack color information, which results in a significant classification error in fine-grained segmentation.

    “However, the pointclouds from these sensors are relatively sparse and lack color information, which results in a significant classification error in fine-grained segmentation [42].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [42] Gao, 2022, Are We Hungry for 3D LiDAR Data for Semantic Segmentation? A Survey of Datasets and Methods, IEEE Trans. Intell. Transp. Syst. source

      Not checked in the free run.

  • The quote only partly states this claim.

    Interpretationreasoning

    Feature-level and late-fusion approaches have been developed to address the limitations of early fusion.

    “Within this reasoning, feature-level and late-fusion approaches have been developed.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Premisenone given

    Very few large-scale (more than 25k labeled samples) semantic segmentation datasets are available for autonomous driving settings, and even fewer take care of the multimodal aspect.

    “Very few large-scale (more than 25k labeled samples) semantic segmentation datasets are available for autonomous driving settings, and even fewer take care of the multimodal aspect of the sensors present in vehicles.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    RADARs can withstand a variety of lighting and weather situations, but due to low resolution, semantic understanding with RADARs is difficult.

    “They can withstand a variety of lighting and weather situations; however, due to their low resolution, semantic understanding with RADARs is difficult.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Working in a fully three-dimensional environment can bring some additional understanding capabilities with respect to the 2D projection given by images.

    “In fact, working in a fully three-dimensional environment can bring some additional understanding capabilities with respect to the 2D projection given by images.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationnone given

    Both symmetrical approaches (same architecture for all modalities) and asymmetrical ones (main modality with others as side information) have been proposed.

    “We also remark that both symmetrical approaches (by using the same architecture for all modalities) and asymmetrical ones (setting a main modality from which the output is computed and by using the others as side information) have been proposed.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Most existing approaches use a projection of the original pointcloud over the color frame and try to find an alignment that can be exploited for the fusion between the crossmodality features.

    “However, most of the existing approaches use a projection of the original pointcloud over the color frame and try to find an alignment that can be exploited for the fusion between the crossmodality features.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Producing a semantically labeled pointcloud is particularly well suited for LiDAR data, which typically have this structure.

    “This approach is particularly well suited for LiDAR data, which typically have this structure.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    For the time being, there is no indication that one fusion scheme is preferable to the others.

    “For the time being, there is no indication that one fusion scheme is preferable to the others.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The fusion of radar data with other approaches is still quite unexplored.

    “Also, the fusion of radar data with other approaches is still quite unexplored.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    ESANet, using a ResNet34 backbone, achieves an mIoU of 80.09 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| ESANet [102] | ResNet34 [16] | 80.09 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [102] Seichter, D., Köhler, M., Lewandowski, B., Wengefeld, T., and Gross, H.M. (June, January 30). Efficient rgb-d semantic segmentation for indoor scene analysis. Proceedings of the 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China. source

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MMSFB-snow, using a ResNet50 backbone, achieves an mIoU of 73.8 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| MMSFB-snow [88] | ResNet50 [16] | 73.8 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [88] Vachmanus, 2021, Multi-Modal Sensor Fusion-Based Semantic Segmentation for Snow Driving Scenarios, IEEE Sens. J. source

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    MDASS, using a VGG16 backbone, achieves an mIoU of 63.13 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| MDASS [108] | VGG16 [15] | 63.13 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [108] Rashed, H., El Sallab, A., Yogamani, S., and ElHelw, M. (2019, January 15–20). Motion and depth augmented semantic segmentation for autonomous navigation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, Long Beach, CA, USA. source

      Not checked in the free run.

    • cites [15] Simonyan, K., and Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv.

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AdapNet, using an AdapNet backbone, achieves an mIoU of 71.72 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| AdapNet [105] | AdapNet [105] | 71.72 |”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [105] Valada, A., Vertens, J., Dhall, A., and Burgard, W. (June, January 29). Adapnet: Adaptive semantic segmentation in adverse environmental conditions. Proceedings of the 2017 IEEE International Conference on Robotics and Automation (ICRA), Singapore. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    RFNet, using a ResNet18 backbone, achieves an mIoU of 69.37 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| RFNet [106] | ResNet18 [16] | 69.37 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [106] Sun, 2019, Rtfnet: Rgb-thermal fusion network for semantic segmentation of urban scenes, IEEE Robot. Autom. Lett. source

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    RSSAWC, using an ICNet backbone, achieves an mIoU of 65.09 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| RSSAWC [77] | ICNet [123] | 65.09 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [77] Pfeuffer, A., and Dietmayer, K. (2019, January 2–5). Robust Semantic Segmentation in Adverse Weather Conditions by means of Sensor Data Fusion. Proceedings of the 22th International Conference on Information Fusion (FUSION), Ottawa, ON, Canada. source

      Not checked in the free run.

    • cites [123] Zhao, H., Qi, X., Shen, X., Shi, J., and Jia, J. (2018, January 8–14). ICNet for Real-Time Semantic Segmentation on High-Resolution Images. Proceedings of the ECCV, Munich, Germany. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    RFBNet, using a ResNet50 backbone, achieves an mIoU of 74.8 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| RFBNet [104] | ResNet50 [16] | 74.8 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [104] Deng, L., Yang, M., Li, T., He, Y., and Wang, C. (2019). RFBNet: Deep multimodal networks with residual fusion blocks for RGB-D semantic segmentation. arXiv.

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    CMFnet, using a VGG16 backbone, achieves an mIoU of 58.97 on the Cityscapes dataset for 2.5D semantic segmentation.

    “| CMFnet [109] | VGG16 [15] | 58.97 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [109] Zhang, 2022, A central multimodal fusion framework for outdoor scene image segmentation, Multimed. Tools Appl. source

      Not checked in the free run.

    • cites [15] Simonyan, K., and Zisserman, A. (2014). Very deep convolutional networks for large-scale image recognition. arXiv.

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DooDLeNet, using a ResNet101 backbone, achieves an mIoU of 57.3 on the MSSSD/MF dataset for RGB + Thermal semantic segmentation.

    “| DooDLeNet [111] | ResNet101 [16] | 57.3 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [111] Frigo, O., Martin-Gaffé, L., and Wacongne, C. (2022, January 19–24). DooDLeNet: Double DeepLab Enhanced Feature Fusion for Thermal-color Semantic Segmentation. Proceedings of the 2022 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LO, USA. source

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AFNet, using a ResNet50 backbone, achieves an mIoU of 54.6 on the MSSSD/MF dataset for RGB + Thermal semantic segmentation.

    “| AFNet [116] | ResNet50 [16] | 54.6 |”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [116] Xu, 2021, Attention fusion network for multi-spectral semantic segmentation, Pattern Recognit. Lett. source

      Not checked in the free run.

    • cites [16] He, K., Zhang, X., Ren, S., and Sun, J. (2016, January 21–26). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. source

      Not checked in the free run.

    • Rests on an assumption: The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.
61 of 61

  • The paper relies on this without support.

    The mIoU values reported by the different cited works are directly comparable, having been obtained under the same evaluation protocol and dataset splits.

    “| Name | Backbone | mIoU |”
1 of 1

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Multimodal fusion strategies can be categorized into early, mid, and late fusion based on fusion strategy.

    “As illustrated in Figure 3, multimodal fusion strategies can be categorized into early fusion, mid fusion, and late fusion based on the fusion strategy [38,39].”

    From Multimodal fusion and vision–language models: A survey for robot vision (2026) · cites this paper as [39]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • SELMA location count differs between text and table

    Minor

    The SELMA description says the dataset covers eight synthetic towns, while its Table 2 entry lists 7C (seven cities).

  • Reference [106] identifies two different architectures

    Minor

    Table 3 assigns reference [106] to RFNet (2020, depth data) and also to RTFNet (2019, thermal data). The two entries cannot both identify the same cited work.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., and Polosukhin, I. (2017, January 4–9). Attention is all you need. Proceedings of the 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
  2. Kolesnikov, A., Dosovitskiy, A., Weissenborn, D., Heigold, G., Uszkoreit, J., Beyer, L., Minderer, M., Dehghani, M., Houlsby, N., and Gelly, S. (2021, January 3–7). An image is worth 16x16 words: Transformers for image recognition at scale. Proceedings of the International Conference on Learning Representations (ICLR), Virtual Event, Austria.
  3. Li, 2020, Lidar for autonomous driving: The principles, challenges, and trends for automotive lidar and perception systems, IEEE Signal Process. Mag., 37, 50
  4. Rosu, R.A., Schütt, P., Quenzel, J., and Behnke, S. (2019). Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv.
  5. Armeni, I., Sax, A., Zamir, A.R., and Savarese, S. (2017). Joint 2D-3D-Semantic Data for Indoor Scene Understanding. arXiv.
  6. Deng, L., Yang, M., Li, T., He, Y., and Wang, C. (2019). RFBNet: Deep multimodal networks with residual fusion blocks for RGB-D semantic segmentation. arXiv.
  7. Iandola, F.N., Moskewicz, M.W., Ashraf, K., Han, S., Dally, W.J., and Keutzer, K. (2016). SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <1 MB model size. arXiv.
  8. Couprie, C., Farabet, C., Najman, L., and LeCun, Y. (2013). Indoor semantic segmentation using depth information. arXiv.

Cited by

Sorted by citations

Related

Authors
Rizzoli, G., Barbato, F., & Zanuttigh, P.
Affiliations
University of Padua (IT)
Corresponding author
Zanuttigh, Pietro
ORCID
Venue
Technologies
Publisher
MDPI AG
Published
2022-07-25
Volume
10
Issue
4
Pages
90
ISSN
2227-7080 (ISSN-L)
DOI
10.3390/technologies10040090
OpenAlex
W4287879563
Semantic Scholar
927717aec038318fba38c319d2371eb0548f8848
Cited by
58
References
129
Citations by year
2026: 5 · 2025: 17 · 2024: 22 · 2023: 13 · 2022: 1
FWCI
4.34
Type
journal article
Language
en
Open access
Licence
Locations
  • Technologies (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Padua Research Archive (University of Padova) (submittedVersion, cc-by)
Funders
SID project “Semantic Segmentation in the Wild”
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Engineering › Automotive Engineering › Autonomous Vehicle Technology and Safety
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
autonomous driving, sensor fusion, LiDAR, depth camera, thermal camera, RGB, deep learning architectures, fusion strategies, perception

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review