Road Anomaly Detection by Partial Image Reconstruction with Segmentation Coupling

Vojíř, Tomáš; Sipka, Tomas; Aljundi, Rahaf; Chumerin, Nikolay; Reino, Daniel Olmeda; Matas, Jiřı́

doi:10.1109/iccv48922.2021.01536

Cited by 54 publications

(31 citation statements)

References 34 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Table 3 presents performance evaluation on SMIYC benchmark. Our method outperforms all previous methods which use image resynthesis [8,9,7], partial image reconstruction [31] or auxiliary data utilization [12].…”

Section: Performance Evaluation On Segmentmeifyoucanmentioning

confidence: 83%

See 1 more Smart Citation

Dense anomaly detection by robust learning on synthetic negative data

Grcić¹,

Bevandić²,

Šegvić³

2021

Preprint

View full text Add to dashboard Cite

Standard machine learning is unable to accommodate inputs which do not belong to the training distribution. The resulting models often give rise to confident incorrect predictions which may lead to devastating consequences. This problem is especially demanding in the context of dense prediction since input images may be only partially anomalous. Previous work has addressed dense anomaly detection by discriminative training on mixed-content images. We extend this approach with synthetic negative patches which simultaneously achieve high inlier likelihood and uniform discriminative prediction. We generate synthetic negatives with normalizing flows due to their outstanding distribution coverage and capability to generate samples at different resolutions. We also propose to detect anomalies according to a principled information-theoretic criterion which can be consistently applied through training and inference. The resulting models set the new state of the art on standard benchmarks and datasets in spite of minimal computational overhead and refraining from auxiliary negative data.

show abstract

Section: Performance Evaluation On Segmentmeifyoucanmentioning

confidence: 83%

“…Still, resynthesis requires significant computational overhead which limits real-world applications. A related approach [31] utilizes a parallel upsampling path for input reconstruction. This improves inference speed with respect to generative resynthesis approaches, but fails to correctly recreate cluttered scenes.…”

Section: Dense Open-set Recognitionmentioning

confidence: 99%

Dense anomaly detection by robust learning on synthetic negative data

Grcić¹,

Bevandić²,

Šegvić³

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Anomalous objects are then identified by the difference between the real and re-synthesized image. For instance, Ohgushi et al [25] and Voijr et al [26] use an autoencoder to re-synthesize the input image where the encoder is part of a semantic segmentation network. The anomaly map is then based on the perceptual loss between the encoder and decoder as well as the entropy loss of the semantic segmentation [25].…”

Section: Related Workmentioning

confidence: 99%

Multimodal Detection of Unknown Objects on Roads for Autonomous Driving

Bogdoll¹,

Enrico²,

Nitsche³

et al. 2022

Preprint

View full text Add to dashboard Cite

Tremendous progress in deep learning over the last years has led towards a future with autonomous vehicles on our roads. Nevertheless, the performance of their perception systems is strongly dependent on the quality of the utilized training data. As these usually only cover a fraction of all object classes an autonomous driving system will face, such systems struggle with handling the unexpected. In order to safely operate on public roads, the identification of objects from unknown classes remains a crucial task. In this paper, we propose a novel pipeline to detect unknown objects. Instead of focusing on a single sensor modality, we make use of lidar and camera data by combining state-of-the art detection models in a sequential manner. We evaluate our approach on the Waymo Open Perception Dataset and point out current research gaps in anomaly detection.

show abstract

“…In earlier works, Kendall et al [14] derived the uncertainty of the segmentation map by Monte Carlo dropout down-sampling, where a higher variance of classes indicates higher uncertainty. Additionally, the uncertainty could be treated as a pixel-level uncertainty score to detect road obstacles, as introduced by [13,15].…”

Section: Related Workmentioning

confidence: 99%

“…Along with the GAN-based approach, the reconstructive approach can detect the road anomaly when it can reproduce the normality of the training data without any auxiliary data of anomalous objects. Vojír et al [15] proposed a recent method that introduces the re-construction module to detect road anomalies based on pixel-wise score maps. Despite promising results, many mentioned methods in GAN-based and reconstructionbased approaches are complicated, power-and time-consuming to retrain complete models and to run on edge devices for real-world applications.…”

Section: Related Workmentioning

confidence: 99%

Effective Free-Driving Region Detection for Mobile Robots by Uncertainty Estimation Using RGB-D Data

Nguyen

et al. 2022

Sensors

View full text Add to dashboard Cite

Accurate segmentation of drivable areas and road obstacles is critical for autonomous mobile robots to navigate safely in indoor and outdoor environments. With the fast advancement of deep learning, mobile robots may now perform autonomous navigation based on what they learned in the learning phase. On the other hand, existing techniques often have low performance when confronted with complex situations since unfamiliar objects are not included in the training dataset. Additionally, the use of a large amount of labeled data is generally essential for training deep neural networks to achieve good performance, which is time-consuming and labor-intensive. Thus, this paper presents a solution to these issues by proposing a self-supervised learning method for the drivable areas and road anomaly segmentation. First, we propose the Automatic Generating Segmentation Label (AGSL) framework, which is an efficient system automatically generating segmentation labels for drivable areas and road anomalies by finding dissimilarities between the input and resynthesized image and localizing obstacles in the disparity map. Then, we train RGB-D datasets with a semantic segmentation network using self-generated ground truth labels derived from our method (AGSL labels) to get the pre-trained model. The results showed that our AGSL achieved high performance in labeling evaluation, and the pre-trained model also obtains certain confidence in real-time segmentation application on mobile robots.

show abstract

Road Anomaly Detection by Partial Image Reconstruction with Segmentation Coupling

Cited by 54 publications

References 34 publications

Dense anomaly detection by robust learning on synthetic negative data

Dense anomaly detection by robust learning on synthetic negative data

Multimodal Detection of Unknown Objects on Roads for Autonomous Driving

Effective Free-Driving Region Detection for Mobile Robots by Uncertainty Estimation Using RGB-D Data

Contact Info

Product

Resources

About