Rapid measurement of steel surface defects is of utmost importance in modern industrial manufacturing. However, challenges arise due to the complex textures and varied morphological patterns of defects. This paper presents YOLO-APEX, an adaptive, lightweight structure for steel surface defect detection. Key innovations include the Multi-Scale Feature Extraction Foundation (MSFEF) module, the Context-Aware Enhancement (CAE) module, and the Adaptive Group Attention Optimization(AGAO) module. The MSFEF captures deep features through horizontal and vertical decomposition at multiple scales, the CAE enhances the understanding of contextual relationships between defects and surrounding textures, and the AGAO integrates a dynamic segmented attention optimization mechanism and a dynamic cardinality grouping strategy to achieve fine-grained, pixel-level feature weighting. Experimental results prove that YOLO-APEX attains a 33.5% mAP 75 on the Steel Surface Defect Detection dataset, outperforming the baseline (29.2% mAP 75 ),and a 43.1% mAP 75 on the NEU-DET dataset,outperforming the baseline (41.8% mAP 75 )while maintaining fewer parameters.

Three-dimensional lane detection is a fundamental yet highly challenging task in autonomous driving, as the presence of interference and blurring in images often impedes accurate detection. To address these challenges, we propose the MRDALane framework, which introduces two novel modules: the Dual Channel Attention Module (DCAM) and the Multi-Resolution Context Augmentation (MRCA). The DCAM utilizes a dual-channel attention mechanism to effectively suppress noise and emphasize salient features, significantly enhancing lane detail capture in complex environments. The MRCA incorporates multiple dilated convolutions with a sawtooth dilation rate design, enabling diverse feature learning across branches and improving robustness across different road conditions. Experimental results demonstrate that MRDALane outperforms state-of-the-art (SOTA) 3D lane detection models, such as LATR and CurveFormer, on both the Apollo and OpenLane datasets. Notably, in robustness experiments under adverse conditions, MRDALane achieved an F1 score improvement of up to 13.66% compared to the LATR model. This comprehensive performance evaluation demonstrates our model’s superior detection accuracy across various challenging scenarios. These advancements provide new insights for future research in 3D lane detection, contributing to the ongoing development of autonomous driving technology and road safety. Our code will be released at https://github.com/Dcelysia/MRDALane.git.

Although 3D point cloud recognition has achieved substantial progress on standard benchmarks, the typical models are vulnerable to point cloud corruptions, leading to security threats in real-world applications. To improve the corruption robustness, various data augmentation methods have been studied, but they are mainly limited to the spatial domain. As the point cloud has low information density and significant spatial redundancy, it is challenging to analyze the effects of corruptions. In this paper, we focus on the frequency domain to observe the underlying structure of point clouds and their corruptions. Through graph Fourier transform (GFT), we observe a correlation between the corruption robustness of point cloud recognition models and their sensitivity to different frequency bands, which is measured by the GFT spectrum of the model’s Jacobian matrix. To reduce the sensitivity and improve the corruption robustness, we propose Frequency Adversarial Training (FAT) that adopts frequency-domain adversarial examples as data augmentation to train robust point cloud recognition models against corruptions. Theoretically, we provide a guarantee of FAT on its out-of-distribution generalization performance. Empirically, we conduct extensive experiments with various network architectures to validate the effectiveness of FAT, which achieves the new state-of-the-art results.

Pose prediction and trajectory forecasting represent pivotal tasks in the realm of autonomous driving, crucially enhancing the planning and decision-making capabilities of self-driving vehicles. However, a prevailing challenge is that many existing algorithms for these tasks necessitate supervised training, demanding substantial human effort and computational resources. To alleviate this resource-intensive burden, this paper introduces an innovative method for predicting future object poses and trajectories in a 3D space, obviating the requirement for manual annotations. The enhanced self-supervised 3D point cloud prediction algorithm proposed in this study incorporates an 3D Action Attention module, augmenting TCNet’s proficiency in extracting vital spatiotemporal and motion information from continuous point cloud range images. Additionally, 3D Octave Convolution is harnessed to mitigate the computational overhead introduced by the 3D Action Attention module, consequently accelerating the model’s inference speed. This advanced self-supervised 3D point cloud prediction algorithm is denoted as TSMNet. TSMNet outperforms the baseline TCNet and several SOTA 3D point cloud prediction models when evaluated on the KITTI Odometry dataset. Moreover, it showcases robust generalization capabilities in unfamiliar environments. Notably, TSMNet can predict future point cloud data for five frames in a mere 33 milliseconds, surpassing the frame rate of typical LiDAR sensors, which typically operate at 10Hz. Furthermore, when integrated with a point cloud clustering and tracking algorithm, the improved self-supervised 3D point cloud prediction algorithm facilitates the extraction of object poses and trajectories. The performance metrics of the point cloud clustering and tracking algorithm attain remarkable levels of accuracy, with a Multiple Object Tracking Accuracy (MOTA) of 86.12% and a Multiple Object Tracking Precision (MOTP) of 91.01% on the KITTI dataset.

3D object detection is an important task in autonomous driving to perceive the surroundings. Despite the excellent performance, the existing 3D detectors lack the robustness to real-world corruptions caused by adverse weathers, sensor noises, etc., provoking concerns about the safety and reliability of autonomous driving systems. To comprehensively and rigorously benchmark the corruption robustness of 3D detectors, in this paper we design 27 types of common corruptions for both LiDAR and camera inputs considering realworld driving scenarios. By synthesizing these corruptions on public datasets, we establish three corruption robustness benchmarks-KITTI-C, nuScenes-C, and Waymo-C. Then, we conduct large-scale experiments on 24 diverse 3D object detection models to evaluate their corruption robustness. Based on the evaluation results, we draw several important findings, including: 1) motion-level corruptions are the most threatening ones that lead to significant performance drop of all models; 2) LiDAR-camerafusion models demonstrate better robustness; 3) camera-only models are extremely vulnerable to image corruptions, showing the indispensability of LiDAR point clouds. We release the benchmarks and codes at https://github.com/thu-ml/3D_Corruptions_AD to be helpful for future studies.

Abstract The use of deep neural networks has revolutionised object tracking tasks, and Siamese trackers have emerged as a prominent technique for this purpose. Existing Siamese trackers use a fixed template or template updating technique, but it is prone to overfitting, lacks the capacity to exploit global temporal sequences, and cannot utilise multi‐layer features. As a result, it is challenging to deal with dramatic appearance changes in complicated scenarios. Siamese trackers also struggle to learn background information, which impairs their discriminative ability. Hence, two transformer‐based modules, the Spatio‐Temporal Fusion (ST) module and the Discriminative Enhancement (DE) module, are proposed to improve the performance of Siamese trackers. The ST module leverages cross‐attention to accumulate global temporal cues and generates an attention matrix with ST similarity to enhance the template's adaptability to changes in target appearance. The DE module associates semantically similar points from the template and search area, thereby generating a learnable discriminative mask to enhance the discriminative ability of the Siamese trackers. In addition, a Multi‐Layer ST module (ST + ML) was constructed, which can be integrated into Siamese trackers based on multi‐layer cross‐correlation for further improvement. The authors evaluate the proposed modules on four public datasets and show comparative performance compared to existing Siamese trackers.

3D perception of objects is critical for many real-world applications, such as autonomous cars and robots. Among them, most state-of-the-art (SOTA) 3D perception systems are based on deep learning models. Recently, the research community found that 3D object classifiers on point cloud based on deep learning are easily fooled by adversarial point cloud craft by attackers. To overcome this, adversarial defenses are considered the most effective ways to improve the robustness of deep learning models, and most adversarial defenses on point cloud are focused on input transformation. However, all previous defense methods decrease the natural accuracy, and the nature of the point cloud classifiers itself has been overlooked. To this end, in this paper, we propose a novel adversarial defense for 3D point cloud classifiers that makes full use of the nature of the point cloud classifiers. Due to the disorder of point cloud, all point cloud classifiers have the property of permutation invariant to the input point cloud. Based on this nature, we design invariant transformations defense (IT-Defense). We show that, even after accounting for obfuscated gradients, our IT-Defense is a resilient defense against SOTA 3D attacks. Moreover, IT-Defense does not hurt clean accuracy compared to previous SOTA 3D defenses. Our code will be available at: https://github.com/cuge1995/IT-Defense.