Let’s Get Dirty: GAN Based Data Augmentation for Camera Lens Soiling Detection in Autonomous Driving
Open accessAbstract
Wide-angle fisheye cameras are commonly used in automated driving for parking and low-speed navigation tasks. Four of such cameras form a surround-view system that provides a complete and detailed view of the vehicle. These cameras are directly exposed to harsh environmental settings and can get soiled very easily by mud, dust, water, frost. Soiling on the camera lens can severely degrade the visual perception algorithms, and a camera cleaning system triggered by a soiling detection algorithm is increasingly being deployed. While adverse weather conditions, such as rain, are getting attention recently, there is only limited work on general soiling. The main reason is the difficulty in collecting a diverse dataset as it is a relatively rare event.We propose a novel GAN based algorithm for generating unseen patterns of soiled images. Additionally, the proposed method automatically provides the corresponding soiling masks eliminating the manual annotation cost. Augmentation of the generated soiled images for training improves the accuracy of soiling detection tasks significantly by 18% demonstrating its usefulness. The manually annotated soiling dataset and the generated augmentation dataset will be made public. We demonstrate the generalization of our fisheye trained GAN model on the Cityscapes dataset. We provide an empirical evaluation of the degradation of the semantic segmentation algorithm with the soiled data.
AI snapshot
Written from the paper’s full text.
TL;DR
The paper proposes GAN-based artificial soiling generation for fisheye camera images, including a CycleGAN baseline pipeline and an end-to-end DirtyGAN that uses a VAE to generate novel soiling patterns and automatic soiling masks. Adding generated soiled images to training improves soiling segmentation accuracy on real test data from 73.95% to 91.71%, and the authors release Dirty WoodScape and Dirty Cityscapes datasets.
Key findings
- Soiling segmentation trained on real images only reached 73.95% mIoU on a real 2K test set; adding 8K generated images raised it to 91.71%, a 17.76% increase. Training solely on generated images reached 47.41%, and real plus standard augmentation reached 78.20%.
- Automatically generated soiling annotations achieved 93.2% mIoU against manual annotation on a random 5% subset, and human inspection judged generated soiling satisfactory for 95% of scenarios.
- Semantic segmentation with DeepLabV3+ degraded under soiling: a clean-trained model scored 56.6% mIoU on clean test data but 34.8% on soiled data, a 21.8% drop; a soiled-trained model scored 52.1% on clean and 48.2% on soiled data.
- On Cityscapes, the same trend appeared: clean-trained model 38.1% clean vs. 26.6% soiled; soiled-trained model 35.5% clean vs. 38.0% soiled.
Methods
- The baseline pipeline uses CycleGAN for clean-to-soiled image translation, a soiling segmentation network trained from weak polygonal WoodScape annotations to estimate masks, Gaussian smoothing, and super-resolution/composition to produce artificially soiled images and automatic annotations.
- DirtyGAN extends this by using a VAE trained on weak soiling annotations to generate novel soiling patterns via convex combinations in latent space, then restricts CycleGAN translation to masked regions.
- Experiments use WoodScape with 10K images and Cityscapes; soiling segmentation uses ResNet50-FCN8 with binary cross-entropy and ADAM at 640×480, while semantic segmentation uses DeepLabV3+ with an 80:20 train/test split.
Limitations
- The paper states the baseline pipeline cannot be expected to work smoothly for water-caused soiling such as raindrops or snow, and that CycleGAN modifies the whole image and the clean-to-soiled generation branch is ill-posed without visual clues for where soiling should appear.
- The paper notes manual polygonal soiling annotation is coarse, time-consuming, costly, and sometimes unreliable, especially at fuzzy boundaries; training solely on generated images also showed limited ability to capture full data diversity.
- Evident from the text: the subjective quality study included expert participants who were biased because they knew the real data, and although dynamic soiling generation is mentioned as possible, the supplementary results present only static soiling.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 5 cited sources checked · 0 supported · 0 flagged
- Findingown data
A model trained on the synthetic soiling data shows a limited degradation to 16.1%.
“A model trained on the synthetic soiling data shows a limited degradation to 16.1%.”
Quote in the paperSubstantial claim - Findingown data
Non-expert participants in the subjective visual study were not able to recognize real images from fakes.
“The non-expert participants were not able to recognize real images from fakes.”
Quote in the paperSubstantial claim - Findingreasoning
Adding generated images along with precisely generated annotations to fully supervised training improves the semantic segmentation output quality.
“If we, however, add the generated images along with the precisely generated annotations, the same fully supervised training exploits these annotations, with a desired effect on the semantic segmentation output quality.”
Quote in the paperSubstantial claim - Findingown data
The image quality of the generated artificial soiling is satisfactory when judged by human inspection for 95% of scenarios.
“In general, we can say that the image quality of the generated artificial soiling is satisfactory when judged by human inspection for 95% of scenarios.”
Quote in the paperSubstantial claim - Findingown data
The Cityscapes dataset results show a similar trend as in the WoodScape experiment.
“The results are presented in Table 3, which, as you can see, show a similar trend as in the Wood-Scape experiment.”
Quote in the paperNot a substantial claim
- Findingown data
Standard data augmentation techniques (flipping, contrast changes) to match the size of training data with generated samples yield a much lesser improvement in accuracy compared to augmentation by generated synthetic samples.
“We also conduct a simple ablation study with standard data augmentation techniques (flipping, contrast changes) to match the size of training data with generated samples and observe a much lesser improvement in accuracy (third row), compared to the scenario with augmentation by generated synthetic samples.”
Quote in the paperSubstantial claim - Findingnone given
The fisheye trained GAN model generalizes to the Cityscapes dataset.
“We demonstrate the generalization of our fisheye trained GAN model on the Cityscapes dataset.”
Quote in the paperSubstantial claim - Findingown data
Segmentation networks can perform segmentation even behind the soiled area.
“Figure 9f shows the capability of segmentation networks to perform segmentation even behind the soiled area.”
Quote in the paperSubstantial claim- Rests on an assumption: In a realistic scenario, annotations are not available for the occluded region of soiled images.
- Findingreasoning
Training the network on artificial images solely causes a significant reduction in accuracy.
“We observe a significant reduction in accuracy when the network is trained on artificial images solely.”
Quote in the paperSubstantial claim - Findingreasoning
The proposed method generates precise annotation compared to coarse manual annotation.
“Thus, in addition to providing novel patterns for augmentation in training, the proposed method also generates precise annotation compared to coarse manual annotation.”
Quote in the paperNot a substantial claim - Findingnone given
The segmentation network M is able to fit the soiling patterns more precisely than the coarse polygonal annotation.
“Even though the annotation is quite coarse, the segmentation network M is able to fit the soiling patterns more precisely.”
Quote in the paperNot a substantial claim - Findingnone given
DirtyGAN can generate similar results as the baseline pipeline in an end-to-end fashion.
“The second algorithm is a novel DirtyGAN network, which can generate similar results as the baseline pipeline in an end-to-end fashion.”
Quote in the paperNot a substantial claim- Rests on an assumption: The possibility to generate random but realistic soiling patterns on camera images is an integral component in examining the degradation of other image processing methods.
- Interpretationreasoning
There is no control over the soiling pattern production process.
“Furthermore, there is also no control over the soiling pattern production process.”
Quote in the paperSubstantial claim - Interpretationnone given
The baseline pipeline cannot be expected to work smoothly for soiling types caused by water (e.g., raindrops, snow) in this specific formulation.
“The biggest one is that it cannot be expected to work smoothly for soiling types caused by water (e.g., raindrops, snow) in this specific formulation.”
Quote in the paperSubstantial claim - Premisecitation
The CycleGAN method modifies the whole image.
“The main problem of the CycleGAN method is that it modifies the whole image.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [34] source
- Interpretationreasoning
The reduction in accuracy when trained on artificial images solely can be attributed to a limited ability to capture the entire data diversity to support the original soiled data.
“It can be attributed to a limited ability to capture the entire data diversity to support the original soiled data.”
Quote in the paperSubstantial claim - Interpretationreasoning
The generation branch from clean to soiled is ill-posed, as there is no visual clue for where the soiling should be produced.
“Secondly, the generation branch”
Quote in the paperSubstantial claim - Premisecitation
CycleGAN does not constrain the image generation to any specific regions and instead regenerates the whole image, affecting all pixels.
“Firstly, CycleGAN does not constrain the image generation to any specific regions and instead regenerates the whole image, affecting all pixels.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [34] source
- Interpretationown data
The generated soiling replicates the real soiling patterns.
“As the generated soiling replicates the real soiling patterns, the network's performance increased to 91.71% on real soiling test data, forming a 17.76% increase in accuracy without the need for costly annotations and real-time soiling scene captures (see the second and fourth row of Table 1).”
Quote in the paperSubstantial claim - Interpretationreasoning
It is not possible to measure the real impact of soiling on the images because a clean version of the same images is needed for a fair comparison.
“Secondly, it is not possible to measure the real impact of soiling on the images, because we need a clean version of the same images for a fair comparison.”
Quote in the paperSubstantial claim - Interpretationreasoning
Expert participants were biased because they knew the real data.
“However, expert participants were biased as they know the real data.”
Quote in the paperNot a substantial claim - Premisecitation
The soiling detection task is necessary for an autonomous driving system because it is used to trigger a camera cleaning system that restores the lens.
“We want to emphasize that the soiling detection task is necessary for an autonomous driving system as it is used to trigger a camera cleaning system that restores the lens [29].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [29] source
- Interpretationown data
The generated soiling data improve soiling detection accuracy without the need for costly annotations and real-time soiling scene captures.
“forming a 17.76% increase in accuracy without the need for costly annotations and real-time soiling scene captures”
Quote in the paperSubstantial claim - Interpretationown data
Soiling can cause severe degradation to a standard visual perception task in autonomous driving.
“This significant drop shows that soiling can cause severe degradation to a standard visual perception task in autonomous driving.”
Quote in the paperSubstantial claim - Interpretationnone given
The proposed method automatically provides the corresponding soiling masks, eliminating the manual annotation cost.
“Additionally, the proposed method automatically provides the corresponding soiling masks eliminating the manual annotation cost.”
Quote in the paperNot a substantial claim
- The quote only partly states this claim. 3
- Classified as probably not a substantial claim. 11
- May combine two claims. 9
The quote only partly states this claim.
Findingown dataThe mIoU accuracy between artificially generated annotation and manual annotation is 93.2% on a random subset of 5% of generated samples.
“We obtain a high accuracy of 93.2%.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Findingnone givenThe generated synthetic soiling patterns are often relatively realistic.
“Besides that, the generated synthetic soiling patterns are often relatively realistic.”
Quote in the paperNot a substantial claimMay combine two claims. The quote only partly states this claim.
Findingown dataTraining on the soiled images shows a 4% accuracy degradation on clean test data compared to the baseline when evaluated on clean images, and the model performs reasonably well on clean data with a few percentage drop in mIoU.
“However, training on the soiled images shows a 4% accuracy degradation on clean test data compared to the baseline when evaluated on clean images.”
- Also stated in: § 4.3. Degradation Effect on Semantic Segmentation · plain
“Note, that model trained on the synthetic data perform reasonably well on the clean data with a few percentage drop in mIoU.”
Quote in the paperSubstantial claim- Also stated in: § 4.3. Degradation Effect on Semantic Segmentation · plain
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Findingnone givenManual annotation can be erroneous, especially at boundaries, and is sometimes worse than artificial annotation based on visual inspection.
“It is to be noted that manual annotation can be subjected, especially at boundaries, and sometimes it is worse than artificial annotation based on visual inspection.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataA DeepLabV3+ segmentation model trained on clean images records 56.6% mIoU on clean test data and 34.8% on soiled data, a performance drop of 21.8%, showing a high level of degradation due to soiling.
“A segmentation model trained on clean images records 56.6% mIoU on clean test data and 34.8% on soiled data, a performance drop of 21.8% compared to clean images.”
- Also stated in: § 4.3. Degradation Effect on Semantic Segmentation · plain
“The baseline model trained on clean images 9a is evaluated on soiled images in 9c showing a high level of degradation due to the soiling.”
Quote in the paperSubstantial claim- Rests on an assumption: Pixel level classification is commonly used in autonomous driving applications.
- Also stated in: § 4.3. Degradation Effect on Semantic Segmentation · plain
May combine two claims. The quote only partly states this claim.
Findingown dataAdding generated soiling data (8k images) to training increased the soiling detection network's performance to 91.71% on real soiling test data, a 17.76% increase in accuracy.
“Augmentation of the generated soiled images for training improves the accuracy of soiling detection tasks significantly by 18% demonstrating its usefulness.”
Quote in the paperSubstantial claimMay combine two claims.
Findingnone givenSegmentation behind the soiled area is less reliable compared to the clean baseline and sensitive to overfitting.
“However, it is less reliable compared to the clean baseline and sensitive to overfitting.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningThe WoodScape Dataset has only coarse polygonal annotation, which is prone to errors, and classical fully supervised training of semantic segmentation for soiling is affected by annotation quality.
“Because only coarse polygonal annotation is available (which is moreover prone to errors), the classical fully supervised training of the semantic segmentation for the soiling is affected by the annotation quality.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataExpert participants were sometimes able to spot a difference in the soiling pattern because it was novel, and they sometimes spotted small artifacts such as blurry texture.
“The expert participants were sometimes able to spot a difference in the soiling pattern because it was novel, and also they sometimes spotted small artifacts, e.g., blurry texture.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Interpretationnone givenThe Dirty WoodScape Dataset can serve as a benchmark for measuring the degradation of off-the-shelf classification algorithms.
“Last but not least, we release a public dataset as a companion to the recently published WoodScape Dataset [33], coined Dirty WoodScape Dataset, which can serve as a benchmark for measuring the degradation of the off-the-shelf classification algorithms.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [33]
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Classified as probably not a substantial claim.
Interpretationnone givenThe mask m obtained by the semantic segmentation network M serves as an automatic annotation of the soiling in the generated image.
“The mask m obtained by the semantic segmentation network M serves as an automatic annotation of the soiling in the generated image.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).
Interpretationnone givenEven disregarding camera cleaning, soiling detection is still needed to increase the uncertainty of vision algorithms in the degraded areas.
“Even disregarding camera cleaning, soiling detection is still needed to increase the uncertainty of vision algorithms in the degraded areas.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S4 (based). May combine two claims.
PremisereasoningHuman annotation of soiling is very time demanding, costly, and not very reliable because precise labeling of soiling on a single image can sometimes be very challenging.
“Human annotation is very time demanding, costly, and not very reliable since the precise labeling of soiling on a single image can sometimes be very challenging, e.g., due to the unclear boundary.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationnone givenIt is possible to use arbitrary images for the final composition once the soiling pattern and mask are obtained.
“Note, it is possible to use arbitrary images for the final composition, once we obtain I s and m.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationSoiling on the lens can be removed by a camera cleaning system that either sprays water or uses a more sophisticated ultrasonic hardware.
“The first significant difference is that soiling on the lens can be removed by a camera cleaning system that either sprays water or uses a more sophisticated ultrasonic hardware [29].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [29] source
May combine two claims.
Interpretationown dataThe soiling model trained on fisheye images generalizes well on the Cityscapes dataset, enabling creation of dirty versions of public datasets.
“We demonstrate that our soiling model trained on fisheye images generalizes well on the Cityscapes dataset, enabling us to create dirty versions of public datasets.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S4 (based).
Premisenone givenThe Gaussian smoothing filter mimics the physical nature of the soiling phenomenon where the edges of the soiling patterns are typically semi-transparent, due to photon scattering.
“We apply the Gaussian smoothing filter because it mimics the physical nature of the soiling phenomenon where the edges of the soiling patterns are typically semi-transparent, due to photon scattering.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
InterpretationreasoningFor the investigation of the soiling impact on further image processing, the background (regions of the image not affected by the soiling) must remain untouched.
“For the investigation of the soiling impact on the further image processing, the background (i.e., regions of the image not affected by the soiling) must remain untouched.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).
Interpretationnone givenSoiling detection is complementary to building segmentation or object detection models robust to soiling without an explicit soiling detection step.
“It is complementary to building segmentation or object detection models robust to soiling without an explicit soiling detection step.”
Quote in the paperNot a substantial claim
The paper relies on this without support.
There is only limited work on general soiling.
“While adverse weather conditions, such as rain, are getting attention recently, there is only limited work on general soiling.”
The paper relies on this without support.
The main reason for the limited work on general soiling is the difficulty in collecting a diverse dataset because soiling is a relatively rare event.
“The main reason is the difficulty in collecting a diverse dataset as it is a relatively rare event.”
The paper relies on this without support.
In a realistic scenario, annotations are not available for the occluded region of soiled images.
“In a realistic scenario, annotations are not available for the occluded region of soiled images.”
The paper relies on this without support.
The possibility to generate random but realistic soiling patterns on camera images is an integral component in examining the degradation of other image processing methods.
“The possibility to generate random but realistic soiling patterns on camera images is an integral component in examining the degradation of other image processing methods.”
The paper relies on this without support.
Pixel level classification is commonly used in autonomous driving applications.
“Pixel level classification is commonly used in autonomous driving applications.”
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · see citing works
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946 · see citing works
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · see citing works
Unpaired image-to-image translation using cycle-consistent adversarial networks
Zhu, J.-Y., Park, T., Isola, P., & Efros, A. A. (2017) · 2017 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.244 · Cited by 23,602 · see citing works
Image-to-Image Translation with Conditional Adversarial Networks
Isola, P., Zhu, J.-Y., Zhou, T., & Efros, A. A. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.632 · Cited by 22,930 · see citing works
Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., & Yuille, A. L. (2018) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2017.2699184 · Cited by 22,392 · see citing works
Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-01234-2_49 · Cited by 20,054 · see citing works
GAN（Generative Adversarial Nets）
淳司, 柴. (2017) · Journal of Japan Society for Fuzzy Theory and Intelligent Informatics · doi:10.3156/jsoft.29.5_177_2 · Cited by 19,268 · see citing works
The Cityscapes Dataset for Semantic Urban Scene Understanding
Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., & Schiele, B. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.350 · Cited by 12,433 · see citing works
Fully convolutional networks for semantic segmentation
Shelhamer, E., Long, J., & Darrell, T. (2017) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2016.2572683 · Cited by 11,898 · see citing works
Self-Attention Generative Adversarial Networks
Zhang, H., Goodfellow, I., Metaxas, D., & Odena, A. (2018) · arXiv · doi:10.48550/arxiv.1805.08318 · Cited by 3,353 · see citing works
Semantic Image Synthesis With Spatially-Adaptive Normalization
Park, T., Liu, M.-Y., Wang, T.-C., & Zhu, J.-Y. (2019) · 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2019.00244 · Cited by 2,974 · see citing works
Fattal, R. (2008) · ACM Transactions on Graphics · doi:10.1145/1360612.1360671 · Cited by 2,141 · see citing works
Unsupervised Image-to-Image Translation NetworksPreprint
Liu, M.-Y., Breuel, T. M., & Kautz, J. (2017) · arXiv (Cornell University) · doi:10.48550/arxiv.1703.00848 · Cited by 2,121 · see citing works
Berman, D., Treibitz, T., & Avidan, S. (2016) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2016.185 · Cited by 1,685 · see citing works
Tutorial on Variational Autoencoders
Doersch, C. (2016) · arXiv · doi:10.48550/arxiv.1606.05908 · Cited by 1,470 · see citing works
The Cityscapes Dataset for Semantic Urban Scene Understanding
Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., & Schiele, B. (2016) · arXiv · doi:10.48550/arxiv.1604.01685 · Cited by 1,174 · see citing works
Progressive Image Deraining Networks: A Better and Simpler Baseline
Ren, D., Zuo, W., Hu, Q., Zhu, P., & Meng, D. (2019) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2019.00406 · Cited by 1,059 · see citing works
Spatial Attentive Single-Image Deraining With a High Quality Real Rain Dataset
Wang, T., Yang, X. S., Xu, K., et al. (2019) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2019.01255 · Cited by 712 · see citing works
Semi Supervised Semantic Segmentation Using Generative Adversarial Network
Souly, N., Spampinato, C., & Shah, M. A. (2017) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2017.606 · Cited by 540 · see citing works
Kim, J., Kim, M., Kang, H.-W., & Lee, K.-H. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1907.10830 · Cited by 503 · see citing works
Heavy Rain Image Restoration: Integrating Physics Model and Conditional Adversarial Learning
Li, R., Cheong, L. F., & Tan, R. T. (2019) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2019.00173 · Cited by 429 · see citing works
Single Image Deraining: A Comprehensive Benchmark Analysis
Li, S., Araujo, I. B., Ren, W., Wang, Z., Tokuda, E. K., Hirata, R., Cesar-Junior, R., Zhang, J., Guo, X., & Cao, X. (2019) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2019.00396 · Cited by 361 · see citing works
WoodScape: A Multi-Task, Multi-Camera Fisheye Dataset for Autonomous Driving
Yogamani, S., Hughes, C., Horgan, J., Sistu, G., Chennupati, S., Uricar, M., Milz, S., Simon, M., Amende, K., Witt, C., et al. (2019) · 2019 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00940 · Cited by 303 · cited 5× · see citing works
“• Public release of an artificial soiling dataset as a companion to the recently published WoodScape Dataset [33], coined Dirty WoodScape Dataset , to encourage further research in this area.”
MethodInfluentialSelf-citationMultimodal Unsupervised Image-to-Image TranslationPreprint
Huang, X., Liu, M.-Y., Belongie, S., & Kautz, J. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1804.04732 · Cited by 259 · see citing works
Deep semantic segmentation for automated driving: Taxonomy, roadmap and challenges
Siam, M., Elkerdawy, S., Jagersand, M., & Yogamani, S. (2017) · 2017 IEEE 20th International Conference on Intelligent Transportation Systems (ITSC) · doi:10.1109/itsc.2017.8317714 · Cited by 178 · see citing works
Self-citationInstaGAN: Instance-aware Image-to-Image TranslationPreprint
Mo, S., Cho, M., & Shin, J. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1812.10889 · Cited by 160 · see citing works
Deep Learning for Seeing Through Window With Raindrops
Quan, Y., Deng, S., Chen, Y., & Ji, H. (2019) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2019.00255 · Cited by 152 · see citing works
Computer vision in automated parking systems: Design, implementation and challenges
Heimberger, M., Horgan, J., Hughes, C., McDonald, J., & Yogamani, S. (2017) · Image and Vision Computing · doi:10.1016/j.imavis.2017.07.002 · Cited by 126 · see citing works
Self-citationPark, T., Liu, M.-Y., Wang, T.-C., & Zhu, J.-Y. (2019) · ACM SIGGRAPH Real-Time Live! · doi:10.1145/3306305.3332370 · Cited by 123 · see citing works
Frame-Consistent Recurrent Video Deraining With Dual-Level Flow
Yang, W., Liu, J., & Feng, J. (2019) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2019.00176 · Cited by 120 · see citing works
RTSeg: Real-Time Semantic Segmentation Comparative Study
Siam, M., Gamal, M., Abdel-Razek, M., Yogamani, S., & Jagersand, M. (2018) · 2018 25th IEEE International Conference on Image Processing (ICIP) · doi:10.1109/icip.2018.8451495 · Cited by 112 · see citing works
Self-citationI Can See Clearly Now: Image Restoration via De-Raining
Porav, H., Bruls, T., & Newman, P. M. (2019) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2019.8793486 · Cited by 103 · see citing works
MODNet: Motion and Appearance based Moving Object Detection Network for Autonomous Driving
Siam, M., Mahgoub, H., Zahran, M., Yogamani, S., Jagersand, M., & El-Sallab, A. (2018) · 2018 21st International Conference on Intelligent Transportation Systems (ITSC) · doi:10.1109/itsc.2018.8569744 · Cited by 95 · see citing works
Self-citationSoilingNet: Soiling Detection on Automotive Surround-View Cameras
Uricar, M., Krizek, P., Sistu, G., & Yogamani, S. (2019) · 2019 IEEE Intelligent Transportation Systems Conference (ITSC) · doi:10.1109/itsc.2019.8917178 · Cited by 55 · see citing works
Self-citationYes, we GAN: Applying adversarial techniques for autonomous driving
Uřičář, M., Křížek, P., Hurych, D., Sobh, I., Yogamani, S. K., & Denny, P. E. (2019) · Electronic Imaging · doi:10.2352/issn.2470-1173.2019.15.avm-048 · Cited by 54 · see citing works
Self-citationOptical Flow in Dense Foggy Scenes Using Semi-Supervised Learning
Yan, W., Sharma, A., & Tan, R. T. (2020) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr42600.2020.01327 · Cited by 49 · see citing works
MODNet: Moving Object Detection Network with Motion and Appearance for Autonomous DrivingPreprint
Siam, M., Mahgoub, H., Zahran, M., et al. (2017) · arXiv (Cornell University) · doi:10.48550/arxiv.1709.04821 · Cited by 42 · see citing works
Ki, S., Sim, H., Choi, J.-S., Kim, S. Y., Seo, S., Kim, S., & Kim, M. (2018) · IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw.2018.00126 · Cited by 37 · see citing works
Desoiling Dataset: Restoring Soiled Areas on Automotive Fisheye Cameras
Uřičář, M., Uličný, J., Sistu, G., et al. (2019) · doi:10.1109/iccvw.2019.00526 · Cited by 33 · see citing works
Self-citationAlletto, S., Carlin, C., Rigazio, L., Ishii, Y., & Tsukizawa, S. (2019) · IEEE International Conference on Computer Vision Workshops · doi:10.1109/iccvw.2019.00286 · Cited by 28 · see citing works
Progressive Image Deraining Networks: A Better and Simpler BaselinePreprint
Ren, D., Zuo, W., Hu, Q., et al. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1901.09221 · Cited by 19 · see citing works
Challenges in Designing Datasets and Validation for Autonomous Driving
Uřičář, M., Hurych, D., Křížek, P., & Yogamani, S. K. (2019) · International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP) · doi:10.5220/0007690700002108 · Cited by 17 · see citing works
Self-citationChallenges in Designing Datasets and Validation for Autonomous Driving
Uřičář, M., Hurych, D., Křížek, P., & Yogamani, S. K. (2019) · International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP) · doi:10.5220/0007690706530659 · Cited by 14 · see citing works
Self-citationMaddu, P., Doherty, W., Sistu, G., et al. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1912.11066 · Cited by 8 · see citing works
Self-citationOptical Flow in Dense Foggy Scenes using Semi-Supervised LearningPreprint
Yan, W., Sharma, A., & Tan, R. T. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2004.01905 · Cited by 3 · see citing works
Flagged first, then by citations
6 of the 52 references couldn’t be matched to a known paper.
Research works printed but not linked: 1
- goodfellow, 2014, Generative adversarial nets, Advances in neural information processing systems
Cited by
Sorted by citations
Perception and sensing for autonomous vehicles under adverse weather conditions: A survey
Zhang, Y., Carballo, A., Yang, H., & Takeda, K. (2023) · ISPRS Journal of Photogrammetry and Remote Sensing · doi:10.1016/j.isprsjprs.2022.12.021 · Cited by 640 · see citing works
Surround-View Fisheye Camera Perception for Automated Driving: Overview, Survey & Challenges
Kumar, V. R., Eising, C., Witt, C., & Yogamani, S. K. (2023) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2023.3235057 · Cited by 121 · see citing works
“As it is difficult to collect soiled data, DirtyGAN [100] proposed to use generative adversarial networks (GANs) to artificially generate different soiling patterns inpainted on real scenes.”
BackgroundSelf-citationBenchmarking the Robustness of LiDAR-Camera Fusion for 3D Object Detection
Yu, K., Tang, T., Xie, H., Lin, Z., Liang, T., Wang, B., Chen, P., Hao, D., Wang, Y., & Liang, X. (2023) · IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw59228.2023.00321 · Cited by 71 · see citing works
“Though there are some works [2,9,30,44,45,47,48,63] explore the robustness of 3D dectors from different perspectives, e.”
BackgroundMethodSynWoodScape: Synthetic Surround-View Fisheye Camera Dataset for Autonomous Driving
Sekkat, A. R., Dupuis, Y., Kumar, V. R., Rashed, H., Yogamani, S. K., Vasseur, P., & Honeiné, P. (2022) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2022.3188106 · Cited by 64 · see citing works
Self-citationWeather and Light Level Classification for Autonomous Driving: Dataset, Baseline and Active Learning
Dhananjaya, M. M., Kumar, V. R., & Yogamani, S. K. (2021) · doi:10.1109/itsc48978.2021.9564689 · Cited by 42 · see citing works
“be effectively used to generate more samples using GAN techniques [21].”
MethodSelf-citationSeeing through obstructions with diffractive cloaking
Shi, Z., Bahat, Y., Baek, S., et al. (2022) · ACM Transactions on Graphics · doi:10.1145/3528223.3530185 · Cited by 34 · see citing works
“The obstructions can be either explicitly detected and suppressed for a downstream task [Uricar et al. 2021], or a light field can be reconstructed from”
BackgroundDeep Learning for Traffic Scene Understanding: A Review
Dolatyabi, P., Regan, J., & Khodayar, M. (2025) · IEEE Access · doi:10.1109/access.2025.3529289 · Cited by 32 · see citing works
Benchmarking Object Detection Robustness against Real-World Corruptions
Liu, J., Wang, Z.-J., Ma, L., Fang, C., Bai, T., Zhang, X., Liu, J., & Chen, Z. (2024) · International Journal of Computer Vision · doi:10.1007/s11263-024-02096-6 · Cited by 31 · see citing works
LiMoSeg: Real-time Bird’s Eye View based LiDAR Motion Segmentation
Mohapatra, S., Hodaei, M., Yogamani, S. K., Milz, S., Gotzig, H., Simón, M., Rashed, H., & Maeder, P. (2022) · International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP) · doi:10.5220/0010866000003124 · Cited by 31 · see citing works
“, 2020), soiling detection (Uricar et al., 2021), (Das et al.”
BackgroundSelf-citationA novel deep learning motivated data augmentation system based on defect segmentation requirements
Niu, S., Peng, Y., Li, B., Qiu, Y., Niu, T., & Li, W. (2023) · Journal of Intelligent Manufacturing · doi:10.1007/s10845-022-02068-y · Cited by 25 · see citing works
All-weather road drivable area segmentation method based on CycleGAN
Chen, J., Depeng, W., Teng, L., et al. (2022) · The Visual Computer · doi:10.1007/s00371-022-02650-8 · Cited by 24 · see citing works
“[50] proposed a new GAN-based algorithm that is able to automatically provide the corresponding pollution masks, eliminating the manual annotation cost.”
BackgroundViT-BEVSeg: A Hierarchical Transformer Network for Monocular Birds-Eye-View Segmentation
Dutta, P., Sistu, G., Yogamani, S. K., Galván, E., & McDonald, J. (2022) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn55064.2022.9891987 · Cited by 19 · see citing works
“solution is to use multiple passive monocular cameras to capture the environment whereby the resulting data is interpreted using computer vision and machine learning techniques for tasks such as object detection [23], distance estimation [24], soiling detection [31, 32], and multi-task learning [3, 28].”
Self-citationAn Online Learning System for Wireless Charging Alignment Using Surround-View Fisheye Cameras
Dahal, A., Kumar, V. R., Yogamani, S. K., & Eising, C. (2022) · IEEE Transactions on Intelligent Transportation Systems · doi:10.1109/tits.2022.3182165 · Cited by 19 · see citing works
“Recently, it has been explored for various tasks such as object detection [12], [13], soiling detection [14], [15], motion segmentation [16], road edge detection [17], weather classification [18], depth”
BackgroundSelf-citationAdversarial Attacks on Multi-task Visual Perception for Autonomous Driving
Sobh, I., Hamed, A., Kumar, V. R., & Yogamani, S. K. (2021) · Journal of Imaging Science and Technology · doi:10.2352/j.imagingsci.technol.2021.65.6.060408 · Cited by 18 · see citing works
“Introduction Autonomous Vehicles are expected to significantly reduce accidents [33], where visual perception systems [5, 6, 7, 8, 11, 32, 36] are in the heart of these vehicles.”
BackgroundSelf-citationLet’s See Clearly: Contaminant Artifact Removal for Moving Cameras
Li, X., Zhang, B., Liao, J., & Sander, P. V. (2021) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00202 · Cited by 16 · see citing works
LiDAR Point Cloud Augmentation for Adverse Conditions Using Conditional Generative Model
Zhang, Y., Ding, M., Yang, H., Niu, Y., Ge, M., Ohtani, K., Zhang, C., & Takeda, K. (2024) · Remote Sensing · doi:10.3390/rs16122247 · Cited by 15 · see citing works
“Acquiring datasets capturing these effects is extremely difficult, leading to the development of models capable of generating soiling effects on cameras and augmenting contamination datasets [22].”
Challenges in Low Infrastructure and Off-Road Automated Driving
Carruth, D. W., Walden, C. T., Goodin, C. T., & Fuller, S. C. (2022) · International Conference on Connected and Autonomous Driving (MetroCAD) · doi:10.1109/metrocad56305.2022.00008 · Cited by 15 · see citing works
Impact of Critical Situations on Autonomous Vehicles and Strategies for Improvement
Beigi, S. A., & Park, B. (2025) · Future Transportation · doi:10.3390/futuretransp5020039 · Cited by 13 · see citing works
L-DIG: A GAN-Based Method for LiDAR Point Cloud Processing under Snow Driving Conditions
Zhang, Y., Ding, M., Yang, H., Niu, Y., Feng, Y., Ohtani, K., & Takeda, K. (2023) · Sensors · doi:10.3390/s23218660 · Cited by 11 · see citing works
“Such an approach has been proven highly effective in rain [13,14], fog [15,16], and snow [17] weather conditions in camera images, plus contaminations on the camera lens [18].”
BackgroundLi, D., Wong, Y. D., Chen, T., Wang, N., & Yuen, K. F. (2024) · Reliability Engineering & System Safety · doi:10.1016/j.ress.2024.110391 · Cited by 10 · see citing works
FisheyePixPro: Self-supervised pretraining using Fisheye images for semantic segmentation
Cheke, R., Sistu, G., Eising, C., van de Ven, P., Kumar, V. R., & Yogamani, S. (2022) · Autonomous Vehicles and Machines · doi:10.2352/ei.2022.34.16.avm-147 · Cited by 11 · see citing works
“The domain of autonomous driving in-Momentumencoder volves object detection [17, 18, 19], soiling detection [20, 21, 22], semantic segmentation [23], weather classification [24, 25], dynamic object detection [26], depth prediction [27, 28, 29, 30, 31], fusion [32], key-point detection and…”
BackgroundSelf-citationUsing deep learning to automate the detection of bird scaring lines on fishing vessels
Acharya, D., Saqib, M., Devine, C., Untiedt, C., Little, L. R., Wang, D., & Tuck, G. N. (2024) · Biological Conservation · doi:10.1016/j.biocon.2024.110713 · Cited by 7 · see citing works
Automotive DNN-Based Object Detection in the Presence of Lens Obstruction and Video Compression
Baris, G., Li, B., Chan, P. H., Avizzano, C. A., & Donzella, V. (2025) · IEEE Access · doi:10.1109/access.2025.3544773 · Cited by 7 · see citing works
“For example, Uricar et al. managed to augment the Cityscapes dataset into Dirty Cityscapes Dataset using a generative adversarial network (GAN) model and an autoencoder through learning real soiled camera data from the WoodScape dataset [40], [41].”
Ensemble-Based Semi-Supervised Learning to Improve Noisy Soiling Annotations in Autonomous Driving
Uřičář, M., Sistu, G., Yahiaoui, L., & Yogamani, S. K. (2021) · IEEE International Conference on Intelligent Transportation Systems (ITSC) · doi:10.1109/itsc48978.2021.9565086 · Cited by 6 · see citing works
“…some work pointed out the problems with manual annotation process and offered artificial generation of training data with pixel-precise annotation [12], [13], there is not much work dedicated to solving the issue of fixing existing imprecise manual annotations or some method to quickly spot the…”
MethodSelf-citationTrained Trajectory based Automated Parking System using Visual SLAM on Surround View CamerasPreprint
Tripathi, N., & Yogamani, S. K. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2001.02161 · Cited by 7 · see citing works
“In general, there is limited work on perception tasks on fisheye cameras but there has been recent progress for tasks such as object detection [14], depth estimation [7], soiling detection [21], trailer detection [1] and multi-task models [16].”
BackgroundSelf-citationExploring Sensor Impact and Architectural Robustness in Adverse Weather on BEV Perception
Kumar, S., Sharma, S., Asghar, R., Mohandas, R., Brophy, T., Sistu, G., Grua, E. M., Donzella, V., & Eising, C. (2025) · IEEE Open Journal of Vehicular Technology · doi:10.1109/ojvt.2025.3621862 · Cited by 5 · see citing works
“Uricar et al. proposed DirtyGAN [51], which introduces synthetic soiling generation to reduce annotation effort and boost detection accuracy, supported by the WoodScape dataset [47].”
Self-citationAnalysis of the Impact of Rain on Perception in Automated Vehicle Applications
Brophy, T., Mullins, D., Parsi, A., Horgan, J., Ward, E., Denny, P., Eising, C., Deegan, B. M., Glavin, M., & Jones, E. (2025) · IEEE Open Journal of Vehicular Technology · doi:10.1109/ojvt.2025.3553718 · Cited by 5 · see citing works
“Water accumulation on the lens surface significantly distorts the image and causes occlusion [6].”
Al-Mohamade, A., Jarraya, S. K., & Alsolami, F. J. (2025) · Journal of Imaging · doi:10.3390/jimaging11040108 · Cited by 4 · see citing works
“Further, in [5], the authors describe how bad weather or muddy roads can lead to the car’s fisheye camera becoming dirty, resulting in parts of the images being lost and affecting the car’s ability to perform certain autonomous driving activities, such as automatic parking.”
Jeong, S., Kim, H., & Cho, Y. (2025) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros60139.2025.11247397 · Cited by 4 · see citing works
“Considering this property, iteratively excluding the points inside the detected ellipse from P s ensures that the remaining points belong to intact regions.”
SpikiLi: A Spiking Simulation of LiDAR based Real-time Object Detection for Autonomous Driving
Mohapatra, S., Mesquida, T., Hodaei, M., Yogamani, S. K., Gotzig, H., & Maeder, P. (2022) · International Conference on Event-based Control, Communication, and Signal Processing · doi:10.1109/ebccsp56922.2022.9845647 · Cited by 4 · see citing works
“Recently, there has been a lot of work on perception related to various tasks such as object detection [1], [2], soiling detection [3], [4], motion segmentation [5], [6], road edge detection [7], [8], weather classification [9], depth prediction [10]–[16], SLAM [17], [18] and in general for multi-task outputs [19]–[25…”
BackgroundSelf-citationEising, C., Horgan, J., & Yogamani, S. K. (2021) · arXiv (Cornell University) · Cited by 4 · see citing works
Self-citationTLGAN: Conditional Style-Based Traffic light Generation with Generative Adversarial Networks
Wang, D., & Ma, X. (2021) · International Conference on High Performance Big Data and Intelligent Systems (HPBD&IS) · doi:10.1109/hpbdis53214.2021.9658470 · Cited by 3 · see citing works
Eberhardt, T. D., Brühl, T., Schwager, R., Sohn, T. S., & Stork, W. (2025) · IEEE Winter Conference on Applications of Computer Vision · doi:10.1109/wacv61041.2025.00507 · Cited by 2 · see citing works
Soiling Builds up Gradually : Soiling Semantic Segmentation
Kim, T., Lee, W., & Hwang, Y. (2024) · International Conference on Control, Automation and Systems (ICCAS) · doi:10.23919/iccas63016.2024.10773253 · Cited by 1 · see citing works
Multi-lane modelling using convolutional neural networks and conditional random fields
Babu, G., Cheke, R., Sistu, G., & Yogamani, S. K. (2022) · Electronic Imaging · doi:10.2352/ei.2022.34.16.avm-148 · Cited by 1 · see citing works
Self-citationThayalan, S., Srıdevı, M., Santhakumar, K., Kiran, B. R., Liu, H., Gauthier, T., & Yogamani, S. K. (2022) · Journal of Electronic Imaging · doi:10.1117/1.jei.32.1.011004 · Cited by 1 · see citing works
Self-citationA Push-Pull Train for Safety of Autonomous Vehicle Software
Wadekar, S. A. (2023) · International Conference on Communication Systems & Networks (COMSNETS) · doi:10.1109/comsnets56262.2023.10041324 · Cited by 1 · see citing works
“Data augmentation has been used to derive images that a soiled sensor may produce [65], or to derive images where the types and/or conditions of lane makers, the number of lanes, weather and light, etc. are varied [57].”
BackgroundMethodEberhardt, T. D., Brühl, T., Schwager, R., Sohn, T. S., & Stork, W. (2024) · IEEE International Conference on Vehicular Electronics and Safety (ICVES) · doi:10.1109/icves61986.2024.10927915 · Cited by 1 · see citing works
“Giventhe challenges in collecting diverse datasets for camera soiling, Uˇriˇc´aˇr et al. [9] propose a GAN-based approach to generate realistic soiling patterns on images from surround-view fisheye cameras.”
Intelligent Systems for Autonomous Mining Operations: Real-Time Robust Road Segmentation
Urrea, C., & Vélez, M. (2025) · Systems · doi:10.3390/systems13090801 · Cited by 1 · see citing works
“…contexts, factors such as adverse weather conditions, illumination variations, and dirt accumulation on Red Green Blue (RGB) camera lenses can generate operational conditions that differ significantly from training domain data, leading to decreased performance of deployed autonomous systems [8].”
Let's See Clearly: Contaminant Artifact Removal for Moving CamerasPreprint
Li, X., Zhang, B., Liao, J., et al. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2104.08852 · Cited by 1 · see citing works
Lee, C., Schätzle, S., Lang, S. A., et al. (2024) · SSRN Electronic Journal · doi:10.2139/ssrn.4844671 · Cited by 1 · see citing works
Ramachandran, S., Sistu, G., McDonald, J., et al. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2107.08246
Self-citationAnalysis of Faster R-CNN network prediction in the presence of lens occlusion and video compressionPreprint
Baris, G., Li, B., Chan, P. H., Avizzano, C. A., & Donzella, V. (2023) · doi:10.36227/techrxiv.23047412.v1
Analysis of Faster R-CNN network prediction in the presence of lens occlusion and video compressionPreprint
Baris, G., Li, B., Chan, P. H., Avizzano, C. A., & Donzella, V. (2023) · doi:10.36227/techrxiv.23047412
Synthetic Soiling Augmentation Using the Diffusion Model for Automotive Camera System
Heo, J., Kim, T., & Hwang, Y. (2024) · International Conference on Platform Technology and Service · doi:10.1109/platcon63925.2024.10830729
“The papers described above and DirtyGAN [21] found the key to obtain realistic data by adopting CycleGAN [27], cycle consistency loss.”
Influence of Water Droplet Contamination for Transparency Segmentation
Knauthe, V., Weitz, P., Pöllabauer, T., Wirth, T., Rak, A., Kuijper, A., & Fellner, D. W. (2025) · Lecture notes in computer science · doi:10.1007/978-981-97-8705-0_17
Beyond the Smudge: Simulating Opaque and Transparent Automotive Camera Lens Soiling
Eberhardt, T. D., Brühl, T., Schwager, R., Sohn, T. S., & Stork, W. (2025) · IEEE Symposia on Computational Intelligence for Energy, Transport and Environmental Sustainability (CIETES) · doi:10.1109/cietes63869.2025.10995030
“(a)Front camera opaque contamination by mud example from the Soiling WoodScape dataset [1] (b)Back camera transparent contamination by dust example from the Soiling WoodScape dataset [1] Fig.”
InfluentialMask Refinement for Vehicle Camera Soiling via Linear Modeling of Border Intensity Transitions
Eberhardt, T. D., Matz, E., Schwager, R., Sohn, T. S., & Stork, W. (2025) · International Conference on Systems, Signals and Image Processing · doi:10.1109/iwssip66997.2025.11151897
“A noteworthy example is the two-stage, ensemble-based semi-supervised learning approach proposed in [3], developed by the same authors who introduced the SoilingNet dataset [2] and subsequent works such as TiledSoilingNet [4] and Let’s Get Dirty [5].”
Robust Few-Shot Semantic Segmentation for Blurred and Occluded Objects in Construction Environments
Salimi, M., Loni, M., Cicchetti, A., & Sirjani, M. (2025) · International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn64981.2025.11227444
“Uricar et al. [36] employed GAN-based data augmentation to generate diverse lens soiling patterns, improving the robustness of autonomous driving systems.”
Realistic Contaminant Image Synthesis using SPADE and Generative Label Synthesis
Byeong-Il, M., Yong-Hyeon, C., Park, J.-M., et al. (2025) · doi:10.1109/icce-asia67487.2025.11263675
“A prominent prior work, DirtyGAN [4], utilized the CycleGAN [6] architecture for contaminant generation.”
Yang, F., Duan, X., Li, F., & Zhang, L. (2026) · Sensors · doi:10.3390/s26113533
“For instance, methods such as Let’s Get Dirty [10] have demonstrated the potential of synthesizing diverse soiling patterns for data augmentation.”
- Authors
- Uricar, M., Sistu, G., Rashed, H., Vobecky, A., Kumar, V. R., Krizek, P., Burger, F., & Yogamani, S.
- Affiliations
- Independent Researcher
- Valeo (France) (FR)
- CTU in Prague
- Valeo
- Venue
- 2021 IEEE Winter Conference on Applications of Computer Vision (WACV)
- Publisher
- IEEE
- Published
- 2021-01
- Pages
- 766-775
- ISSN
- 2472-6737 (ISSN-L)
- OpenAlex
- W3106024689
- Semantic Scholar
- 19a6d112e4750078102c961f47ede4a63282cffe
- Cited by
- 51
- References
- 52
- Citations by year
- 2026: 1 · 2025: 14 · 2024: 8 · 2023: 8 · 2022: 11 · 2021: 8 · 2020: 1
- FWCI
- 3.04
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- IEEE Winter Conference on Applications of Computer Vision (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- České Vysoké Učení Technické v Praze
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Generative Adversarial Networks and Image Synthesis
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image Enhancement Techniques
- Keywords
- generative adversarial networks, data augmentation, fisheye camera, autonomous driving, semantic segmentation, synthetic image generation
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.
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
