CoChat
Sign Up
Cited by 303Advanced Neural Network ApplicationsAdvanced Vision and ImagingAutonomous Vehicle Technology and Safety

WoodScape: A Multi-Task, Multi-Camera Fisheye Dataset for Autonomous Driving

Open access
Yogamani, Senthil, Hughes, Ciaran, Horgan, Jonathan, Sistu, Ganesh, Chennupati, Sumanth, Uricar, Michal, Milz, Stefan, Simon, Martin +9 more·2019·2019 IEEE/CVF International Conference on Computer Vision (ICCV)
DOI 10.1109/iccv.2019.00940
Open this paper in CoChat — free accountOpen PDF

Abstract

Fisheye cameras are commonly employed for obtaining a large field of view in surveillance, augmented reality and in particular automotive applications. In spite of their prevalence, there are few public datasets for detailed evaluation of computer vision algorithms on fisheye images. We release the first extensive fisheye automotive dataset, WoodScape, named after Robert Wood who invented the fisheye camera in 1906. WoodScape comprises of four surround view cameras and nine tasks including segmentation, depth estimation, 3D bounding box detection and soiling detection. Semantic annotation of 40 classes at the instance level is provided for over 10,000 images and annotation for other tasks are provided for over 100,000 images. With WoodScape, we would like to encourage the community to adapt computer vision models for fisheye camera instead of using naive rectification.

AI snapshot

Written from the paper’s full text.

TL;DR

WoodScape is introduced as a multi-task, multi-camera fisheye dataset for autonomous driving, with four 190° horizontal-FOV cameras, LiDAR, GNSS/IMU, and annotations for nine tasks. It provides baseline results across segmentation, detection, depth, motion, soiling, odometry/SLAM, and 3D boxes, while arguing for native fisheye models rather than undistortion.

Key findings

  • The dataset contains over 10,000 images with instance-level semantic annotation, four 1MPx RGB fisheye cameras covering 360°, one Velodyne HDL-64E LiDAR at 20 Hz, GNSS/IMU, and labels for nine tasks including semantic segmentation, 2D/3D detection, depth, motion segmentation, soiling detection, visual odometry/SLAM, and end-to-end driving.
  • Baseline results include ENet segmentation IoU 51.4, Faster R-CNN 2D detection mAP 31 at IoU ≥ 0.5, ResNet10 soiling category precision 84.5%, Eigen depth RMSE 7.7, MODNet motion segmentation IoU 45, visual odometry translation <5 mm 51% and rotation <0.1° 71%, and LSD-SLAM relocalization 61%.
  • For 3D bounding box detection, Complex-YOLO trained with the proposed S_srt metric achieved AP 62.46%, AOS 88.43%, and runtime 1 ms, versus 3D-IoU training with AP 64.38%, AOS 85.60%, and runtime 95 ms, improving training time by 95×.
  • A pre-trained person detector achieved mAP 12 on fisheye images versus 45 for a model trained on WoodScape, which the authors use to argue that fisheye geometry must be handled explicitly.

Methods

  • The dataset was captured from three geographical locations (USA, Europe, China), with highway, urban, and parking scenarios, saloon and sports utility vehicles, and intrinsic/extrinsic calibration plus timestamp synchronization.
  • Annotation covers 40 semantic classes, 5k soiling images, 10k 3D box frames, 400k depth frames, 10k motion segmentation frames, 50 video sequences with over 100k frames for VO/SLAM, 500 end-to-end driving videos, and 10k synthetic frames.
  • Data splitting is 6:1:3 training/validation/testing, with five splits and task-specific stratified splits; sampling uses a wrapper method based on classifier accuracy plus a filter based on image similarity. Baselines use ENet, Faster R-CNN with ResNet101, ResNet10, Eigen, MODNet, LSD-SLAM, and Complex-YOLO.

Limitations

  • The paper states that 2D bounding box detection mAP of 31 is significantly lower than on other datasets and that incorporating fisheye geometry explicitly remains an open research problem.
  • The paper states that spherical models do not accurately fit fisheye lenses and that adapting CNNs to fisheye distortion is an open problem; it also notes that simultaneous annotation is provided for various tasks but with exceptions due to practical limitations of optimal dataset design.
  • The paper states that more than one third of the dataset is recorded in Europe and is GDPR-sensitive due to visible faces and license plates, so it follows a license-agreement approach rather than blurring or GAN retargeting; it also states that end-to-end steering/braking is not currently mature for deployment.

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

Integrity signals

Registry

Crossref · 10.1109/iccv.2019.00940

Cites

69 of 69 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

13 of 34 cited sources checked · 0 supported · 1 flagged · accuracy not yet measured

  • Findingown data

    Running the semi-supervised approach from [33] with the Eigen [14] model as baseline on the WoodScape dataset obtained an RMSE of 7.7 for monocular depth estimation.

    “We run the semi-supervised approach from [33] using the model proposed by Eigen [14] as baseline on our much larger dataset and obtained an RMSE (Root Mean Square Error) value of 7.7.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [33] source
    • cites [14] source

      The citation in the sentence attributes the semi-supervised approach to this source; the RMSE is presented as the citing authors’ own result. Section 2 (pp. 2–3) instead describes training a network on images paired with human steering commands to predict steering, not a semi-supervised approach to monocular depth estimation. (LLM, conf=0.86)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Findingown data

    Soiling Detection with ResNet10 achieves a Category (%) of 84.5 on the WoodScape dataset.

    “| Soiling Detection | ResNet10 [20] | Category (%) | 84.5 |”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] source

      The paper’s discussion of classification methods (III. Vision Algorithm Building Blocks, p. 4) and its table of established applications (p. 5) do not report a soiling-detection result. Searches for WoodScape, ResNet10, and 84.5 found no passage supporting the attributed dataset, model, or score. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
  • Study design and conductcitation

    Task-specific splits are provided because class imbalance may be an issue depending on the task.

    “Depending on the particular task (see Section 4, for the full list), the class imbalance may be an issue [19], therefore, task-specific splits are also provided.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [19] source

      Section 5.1 (p. 12) identifies a class-imbalance problem: traffic lights occur less often than light poles, and the authors suggest better class-balancing strategies. Section 5.2 (p. 12) describes separate task benchmarks and dataset splits, but does not say those splits were provided because of class imbalance. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Findingown data

    Visual Odometry with ResNet50 achieves a Translation (<5mm) score of 51 on the WoodScape dataset.

    “| Visual Odometry | ResNet50 [20] | Translation (<5mm) Rotation (<0.1°) | 51 71 |”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] source

      The discussion of odometry in 'Towards a generic scalable algorithm framework' (p. 6) concerns its use in segmenting moving objects, not a performance evaluation. Searches for WoodScape, ResNet50, and the claimed translation threshold and score found no passage in our copy reporting that result. (LLM, conf=0.77)

      Experimental — accuracy not yet measured
  • Findingown data

    Visual Odometry with ResNet50 achieves a Rotation (<0.1°) score of 71 on the WoodScape dataset.

    “| Visual Odometry | ResNet50 [20] | Translation (<5mm) Rotation (<0.1°) | 51 71 |”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] source

      The Motion Estimation section (p. 3) describes tracking and matching image pixels, but reports no visual-odometry benchmark score. Searches for WoodScape, ResNet50, and the stated rotation and translation thresholds found no passage supporting the score of 71. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
  • May combine two claims. 36
  • Classified as probably not a substantial claim. 19

  • May combine two claims.

    Premisenone given

    There are very few public fisheye datasets and none of them provide semantic segmentation annotation.

    “There are very few public fisheye datasets and none of them provide semantic segmentation annotation.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [7] source

      The RobotCar Platform (p. 1) describes fisheye cameras, and the paper says the dataset is available for download. But the Introduction (p. 1) does not assess how many public fisheye datasets exist, and searches for semantic-segmentation annotations found no passage saying that none provide them. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • cites [9] source
    • cites [37]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisecitation

    The authors of the EuroCity persons dataset argued that any anonymization measure will introduce a bias, and released their dataset with original data and a license agreement enforcing GDPR adherence; WoodScape will follow a similar approach.

    “In the recent EuroCity persons dataset [5], the authors argued that any anonymization measure will introduce a bias.Thus they released their dataset with original data and a license agreement which enforces the user to strictly adhere to GDPR.We will follow a similar approach.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [5] source

      Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

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

    Premisecitation

    Bojarski et al. demonstrated end-to-end learning for steering [2] and recently it was applied to fisheye cameras [54].

    “Bojarski et al. demonstrated end-to-end learning [2] for steering and recently it was applied to fisheye cameras [54].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [2] source
    • cites [54] source

      The Abstract and Introduction discuss semantic segmentation, but neither addresses end-to-end steering with fisheye cameras. Searches for fisheye steering and Bojarski found no relevant passage in our copy. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

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

    Premisenone given

    There are several popular synthetic datasets like SYNTHIA [44] and CARLA [13].

    “There are several popular synthetic datasets like SYNTHIA [44] and CARLA [13].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [44] source

      The citation is attached to SYNTHIA, but the passages from “Selecting Representative Data Sets” (page unavailable) discuss representative-set selection and SMOTE synthetic instances, not SYNTHIA. Searches for SYNTHIA and CARLA found no passage identifying either dataset or addressing its popularity. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [13]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisereasoning

    Semantic segmentation networks for autonomous driving have been successfully trained directly on fisheye images, but due to absence of fisheye datasets, they use artificially warped images of Cityscapes for training and test on fisheye images; artificial images cannot increase the originally captured FOV.

    “Semantic segmentation networks for autonomous driving [47] have been successfully trained directly on fisheye images in [12,45].Due to absence of fisheye datasets, they make use of artificially warped images of Cityscapes for training and testing was performed on fisheye images.However, the artificial images cannot increase the originally captured FOV.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [47] source

      Split into 2 atoms — 0 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 2 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [12] source
    • cites [45] source
    • Rests on an assumption: Artificially warped images cannot increase the originally captured field of view.

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

    Premisecitation

    Motion segmentation has been explored in [49,57,48] for narrow angle cameras.

    “This has been explored in [49,57,48] for narrow angle cameras.”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites [49]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [57]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [48] source

      The Introduction (p. 1) describes semantic segmentation for automated driving, and the Auxiliary Task Learning section (p. 3) illustrates optical flow as a possible task. Searches for motion segmentation and narrow-angle cameras found no passage saying this paper explored motion segmentation with such cameras. (LLM, conf=0.73)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Study design and conductcitation

    The WoodScape dataset supports correct hypothesis evaluation, and therefore multiple splits are provided (5 in total).

    “The dataset supports correct hypothesis evaluation [55], therefore multiple splits are provided (5 in total).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [55] source

      The section on large annotated datasets discusses CamVid, Cityscapes and other datasets, but not WoodScape. The benchmarking section describes an evaluation on CamVid; neither it nor the other passages found addresses WoodScape’s support for hypothesis evaluation or its five splits (searched: WoodScape, hypothesis evaluation, dataset splits). (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Premisecitation

    3D IoU is generally used to evaluate 3D bounding box predictions.

    “In general, 3D IoU [18] is used to evaluate 3D bounding box predictions”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [18] source

      Section 4.6, “A note on accuracy” (p. 11), discusses detection and localisation accuracy for automated parking, but does not identify 3D IoU as an evaluation metric. Searches for 3D IoU, intersection over union, and bounding-box evaluation found no passage addressing the claim. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

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

    Findingnone given

    Higher orders than fourth provide no additional accuracy for the fisheye intrinsic model.

    “In our experience, higher orders provide no additional accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    WoodScape provides model parameters for a generic fisheye intrinsic calibration that is independent of any specific projection model and does not require distortion correction.

    “In WoodScape, we provide model parameters for a more generic fisheye intrinsic calibration that is independent of any specific projection model, and does not require the added step of distortion correction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    WoodScape introduces a novel soiling detection task and releases the first dataset of its kind.

    “Introduction of a novel soiling detection task and release of first dataset of its kind.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    WoodScape contains four cameras covering 360° accompanied by a HD laser scanner, IMU and GNSS.

    “It contains four cameras covering 360 ° accompanied by a HD laser scanner, IMU and GNSS.”
    • Also stated in: § 1. Introduction · plain
      “WoodScape is a comprehensive dataset for 360 ° sensing around a vehicle using the four fisheye cameras shown in Figure 2.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    WoodScape is a four-camera nine-task dataset designed to encourage unified multi-task and multi-camera models.

    “Four-camera nine-task dataset designed to encourage unified multi-task and multi-camera models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For low incident angles, the difference between the fourth order polynomial and UCM is about four pixels, and for eUCM about one pixel; for larger incident angles, these models are less precise.

    “The plot indicates that the difference to the original fourth order polynomial is about four pixels for UCM and one pixel for eUCM for low incident angles.For larger incident angles, these models are less precise.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    WoodScape is the first fisheye image dataset dedicated to autonomous driving.

    “We introduce WoodScape, the first fisheye image dataset dedicated to autonomous driving.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The majority of WoodScape data was obtained from saloon vehicles, with a significant subset from a sports utility vehicle, ensuring a strong mix in sensor mechanical configurations.

    “While the majority of data was obtained from saloon vehicles there is a significant subset from a sports utility vehicle ensuring a strong mix in sensor mechanical configurations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    More than one third of the WoodScape dataset is recorded in Europe and is therefore GDPR sensitive due to visible faces of pedestrians and license plates.

    “More than one third of our dataset is recorded in Europe and is therefore GDPR sensitive due to visible faces of pedestrians and license plates.”
    Quote in the paperSubstantial claim

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

    Findingnone given

    WoodScape driving scenarios are divided across highway, urban driving and parking use cases.

    “Driving scenarios are divided across the highway, urban driving and parking use cases.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The video sequences can also be used for Visual SLAM, where the focus is on relocalization of a mapped trajectory and the metric is the same as for visual odometry.

    “The video sequences can also be used for Visual SLAM where we focus on relocalization of a mapped trajectory and the metric is same as VO.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    For 3D bounding box detection with Complex-YOLO, training optimized on S_srt achieves similar average precision (3D-IoU), better angle orientation similarity (AOS), and much faster computation time compared to training optimized on 3D-IoU.

    “Even though this comparison uses 3D-IoU, we achieve similar performance for average precision (3D-IoU), with better angle orientation similarity (AOS) and much faster computation time.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    WoodScape provides a set of synthetic data using accurate models of the real cameras, enabling investigations of additional tasks.

    “The dataset also provides a set of synthetic data using accurate models of the real cameras, enabling investigations of additional tasks.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    The WoodScape dataset has added end-to-end steering and braking tasks to encourage modular end-to-end architectures and to explore optimized perception for the control task.

    “Thus we have added end-to-end steering and braking tasks to encourage modular end-to-end architectures and to explore optimized perception for the control task.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The WoodScape dataset provides 50 video sequences comprising over 100,000 frames with ground truth.

    “We provide 50 video sequences comprising of over 100k frames with ground truth.”
    Quote in the paperNot a substantial claim

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

    Findingnone given

    WoodScape images are provided at 1MPx 24-bit resolution and videos are uncompressed at 30fps ranging in duration from 30s to 120s.

    “Images are provided at 1MPx 24-bit resolution and videos are uncompressed at 30fps ranging in duration from 30s to 120s.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    WoodScape includes 4x 1MPx RGB fisheye cameras (190° horizontal FOV), 1x LiDAR rotating at 20Hz (Velodyne HDL-64E), 1x GNSS/IMU (NovAtel Propak6 & SPAN-IGM-A1), 1x GNSS Positioning with SPS (Garmin 18x), and odometry signals from the vehicle bus.

    “• 4x 1MPx RGB fisheye cameras (190 • horizontal FOV) • 1x LiDAR rotating at 20Hz (Velodyne HDL-64E) • 1x GNSS/IMU (NovAtel Propak6 & SPAN-IGM-A1) • 1x GNSS Positioning with SPS (Garmin 18x) • Odometry signals from the vehicle bus.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    WoodScape provides intrinsic and extrinsic calibrations for all sensors as well as timestamp files to allow synchronization.

    “Intrinsic and extrinsic calibrations are provided for all sensors as well as timestamp files to allow synchronization of the data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    WoodScape provides 40 classes and 10,000 frames for semantic segmentation.

    “| Segmentation | Classes Frames | 8 400 | 30 5k | 66 25k | -- | 25 140k | 40 5.7k | 40 10k |”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    The baseline 2D object detection model achieves a mAP score of 31, which is significantly less than the accuracy achieved in other datasets.

    “We achieve a mAP score of 31 which is significantly less than the accuracy achieved in other datasets.”
    Quote in the paperSubstantial claim

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

    Findingown data

    The WoodScape dataset also provides previous and next images for exploring multi-stream models like MOD-Net [49].

    “We also provide previous and next images for exploring multi-stream models like MOD-Net [49].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [49] source

  • May combine two claims.

    Findingown data

    WoodScape provides data for nine tasks in total, more than is typical, including completely novel tasks such as soiled lens detection.

    “In addition to providing fisheye data, we provide data for many more tasks than is typical (nine in total), providing completely novel tasks such as soiled lens detection.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The WoodScape dataset provides 3D bounding box annotation for 10,000 frames with three classes: pedestrian, vehicle, and cyclist.

    “3D box annotation is provided for 10k frames with 3 classes namely 'pedestrian', 'vehicle' and 'cyclist'.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The WoodScape dataset provides more than 100,000 images from all four cameras (totaling 400,000) with ground truth provided by LiDAR for monocular depth estimation.

    “We provide more than 100k images of all four cameras (totaling 400k) using ground truth provided by LiDAR.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    WoodScape provides 7 classes and 10,000 frames for 2D bounding box detection.

    “| 2D Bounding Box 1 | Classes Frames | 3 15k | -- | -- | -- | -- | 10 5.7k | 7 10k |”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Fisheye is particularly beneficial to automotive low speed manoeuvring scenarios such as parking where accurate full coverage near field sensing can be achieved with just four cameras.

    “Fisheye is particularly beneficial to automotive low speed manoeuvring scenarios such as parking [21] where accurate full coverage near field sensing can be achieved with just four cameras.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [21]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Premisenone given

    Spherical models do not provide an accurate fit for fisheye lenses and it is an open problem.

    “However, spherical models do not provide an accurate fit for fisheye lenses and it is an open problem.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [52] source
    • cites [9] source

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

    Interpretationreasoning

    Because of fundamental limitations of undistortion, adapting the algorithm by incorporating the fisheye model could be an optimal solution.

    “Because of fundamental limitations of undistortion, an alternate approach of adapting the algorithm incorporating fisheye model discussed in previous section could be an optimal solution.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisereasoning

    Images from fisheye cameras display severe distortion due to complex projection geometry.

    “This advantage comes with some drawbacks in the significantly more complex projection geometry that fisheye cameras exhibit.That is, images from fisheye cameras display severe distortion.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    Rectilinear correction of fisheye images leads to a loss of FOV because the FOV is greater than 180° and there are rays incident from behind the camera, making a complete mapping to a rectilinear viewport impossible.

    “Firstly, the FOV is greater than 180 • , hence there are rays incident from behind the camera and it is not possible to establish a complete mapping to a rectilinear viewport.This leads to a loss of FOV, this is seen via the missing yellow pillars in the corrected image.”
    Quote in the paperSubstantial claim

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

    Premisereasoning

    Fisheye cameras can observe the full surrounding of a vehicle with a minimal number of sensors, with just four cameras typically being required for full 360° coverage.

    “Given their extremely wide field of view, they can observe the full surrounding of a vehicle with a minimal number of sensors, with just four cameras typi- cally being required for full 360 • coverage (Figure 2).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Standard computer vision models do not generalize easily to fisheye cameras because of large non-linear distortion; for example, translation invariance is lost for a standard CNN.

    “Standard computer vision models do not generalize easily to fisheye cameras because of large non-linear distortion.For example, translation invariance is lost for a standard convolutional neural net (CNN).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    Existing datasets are primarily designed to facilitate task-specific learning and they do not provide simultaneous annotation for all tasks.

    “Existing datasets are primarily designed to facilitate task-specific learning and they don't provide simultaneous annotation for all the tasks.”
    Quote in the paperNot a substantial claim

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

    Study design and conductnone given

    Full control of the splitting mechanism is provided, allowing for each class to be represented equally within each split (stratified sampling).

    “Full control of the splitting mechanism is provided allowing for each class to be represented equally within each split (i.e.stratified sampling).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning

    Surround view cameras are usually directly exposed to adverse environmental conditions and thus prone to becoming soiled or having water drops form on the lens.

    “Unlike the front camera which is behind the windshield, the surround view cameras are usually directly exposed to the adverse environmental conditions, and thus prone to becoming soiled or water drops forming on the lens.”
    Quote in the paperSubstantial claim

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

    Premisenone given

    The classes to detect for soiling are {opaque, transparent}; opaque soiling typically arises from mud and dust, and transparent soiling from water and ice, though water can produce 'opaque' regions in practice.

    “The classes to detect are {opaque, transparent}.Typically, opaque soiling arises from mud and dust (Figure 8 right image), and transparent soiling arises from water and ice (Figure 8 left image).However, in practice it is common to see water producing”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Opaque soiling typically arises from mud and dust, and transparent soiling from water and ice.
    • Rests on an assumption: Water can produce 'opaque' regions in the camera image.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    High-quality data in WoodScape is ensured via quality checks at all stages of data collection, and annotation data undergoes rigorous quality assurance by highly skilled reviewers.

    “High-quality data is ensured via quality checks at all stages of the data collection process.Annotation data undergoes a rigorous quality assurance by highly skilled reviewers.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    It is necessary to incorporate the fisheye geometry explicitly for 2D object detection on fisheye images, which is an open research problem.

    “We observe that it is necessary to incorporate the fisheye geometry explicitly, which is an open research problem.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    Collecting frames at 30fps, particularly at low speeds, leads to significant image overlap, making an effective sampling strategy critical.

    “Collecting frames at a frame rate of 30fps, particularly at low speeds, ultimately leads to significant image overlap, therefore, having an effective sampling strategy to distill the dataset is critical.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    The task of soiling detection was first defined in [56].

    “The task of soiling detection was to our best knowledge first defined in [56].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [56] source

  • Classified as probably not a substantial claim.

    Premisecitation

    Careful inspection of training sets for outliers improves the robustness of deep neural networks, especially with regards to adversarial examples.

    “However, at the same time, it was shown that careful inspection of the training sets for outliers improves the robustness of deep neural networks [36], especially with regards to the adversarial examples.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [36] source

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

    Premisecitation

    Visual Odometry (VO) is necessary for creating a map from the objects detected.

    “Visual Odometry (VO) is necessary for creating a map from the objects detected [38].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [38] source
50 of 50

  • The paper relies on this without support.

    Surround view systems have at least four cameras rigidly connected to the body of the car.

    “Surround view systems have at least four cameras rigidly connected to the body of the car.”

  • The paper relies on this without support.

    Artificially warped images cannot increase the originally captured field of view.

    “However, the artificial images cannot increase the originally captured FOV.”

  • The paper relies on this without support.

    Opaque soiling typically arises from mud and dust, and transparent soiling from water and ice.

    “Typically, opaque soiling arises from mud and dust (Figure 8 right image), and transparent soiling arises from water and ice (Figure 8 left image).”

  • The paper relies on this without support.

    Water can produce 'opaque' regions in the camera image.

    “However, in practice it is common to see water producing”

  • The paper relies on this without support.

    3D IoU has drawbacks, especially for rotated objects.

    “In general, 3D IoU [18] is used to evaluate 3D bounding box predictions, but there are drawbacks, especially for rotated objects.”

  • The paper relies on this without support.

    Two boxes can reach a good 3D IoU score while overlapping in total with an opposite heading.

    “Two boxes can reach a good 3D IoU score, while overlapping in total with an opposite heading.”

  • The paper relies on this without support.

    Exact calculation in 3D space is time-consuming.

    “Additionally, an exact calculation in 3D space is a time consuming task.”

  • The paper relies on this without support.

    Motion is a strong cue in automotive due to ego-motion of cameras and dynamic objects are critical interacting agents.

    “In automotive, motion is a strong cue due to ego-motion of the cameras on the moving vehicle and dynamic objects around the vehicle are the critical interacting agents.”

  • The paper relies on this without support.

    Synthetic data is crucial for autonomous driving.

    “Synthetic data is crucial for autonomous driving for many reasons.”

  • The paper relies on this without support.

    There are legal restrictions like recording videos of a child.

    “Secondly, there are legal restrictions like recording videos of a child.”

  • The paper relies on this without support.

    Synthetic data is the only way to obtain dense depth and optical flow annotation.

    “Finally, synthetic data is the only way to obtain dense depth and optical flow annotation.”
11 of 11

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?

  • Image restoration using inpainting or generative methods carries risks of hallucinated visual cues that may mislead the perception system.

    “Other works explored image restoration using inpainting or generative methods [3,4], though such approaches carry risks of hallucinated visual cues that may mislead the perception system.”

    From AI-Based Image Quality and Lens Defect Analysis in Autonomous Driving: A Framework with U-Net-Based Soiling Detection (2025) · cites this paper as [4]

    Section 2.2 discusses geometric undistortion of fisheye images, not restoration by inpainting or generative methods. The GDPR challenges section mentions GAN-generated face replacement for privacy, but does not address hallucinated visual cues or their effects on perception. Searches for inpainting, restoration, and hallucinated cues found no passage supporting the attributed point. (LLM, conf=0.75)

    Experimental — accuracy not yet measured

Cites

Flagged first, then by citations

9 of the 78 references couldn’t be matched to a known paper.

Research works printed but not linked: 2

  1. kumar, 2018, Near-field depth estimation using monocular fisheye camera: A semi-supervised learning approach using sparse LiDAR data, CVPR Workshop
  2. caruso, 2015, Largescale direct SLAM for omnidirectional cameras, Proceedinsg of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 141

Cited by

Sorted by citations

Related

Authors
Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Chennupati, S., Uricar, M., Milz, S., Simon, M., Amende, K., Witt, C., Rashed, H., Nayak, S., Mansoor, S., Varley, P., Perrotton, X., Odea, D., & Perez, P.
Affiliations
Valeo (France) (FR)
ORCID
Venue
2019 IEEE/CVF International Conference on Computer Vision (ICCV)
Publisher
IEEE
Published
2019-10
Pages
9307-9317
DOI
10.1109/iccv.2019.00940
OpenAlex
W2943449463
Semantic Scholar
57de1a6cc6035d4ac6305121367ad3c6d3fd5104
Cited by
303
References
78
Citations by year
2026: 19 · 2025: 50 · 2024: 55 · 2023: 70 · 2022: 44 · 2021: 39 · 2020: 18 · 2019: 8
FWCI
8.80
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE/CVF International Conference on Computer Vision (ICCV) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Vision and Imaging
  • Physical Sciences › Engineering › Automotive Engineering › Autonomous Vehicle Technology and Safety
Keywords
fisheye camera, surround-view cameras, autonomous driving, semantic segmentation, depth estimation, soiling detection

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 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Cited sources are checked by gpt-6-sol.

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