FORTRESS: Feature Optimization and Robustness Techniques for 3D Object Detection Systems

Caixin Kang, Xinning Zhou, Chengyang Ying, Wentao Shang, Xingxing Wei, Yinpeng Dong, Hang Su

Keywords: Robustness, 3D Object Detection, Autonomous Driving, Corner Cases

Subject: 3D object detection and scene understanding

Confirmation: I have read and agree with the submission policies of ECCV 2024 and the W-CODA Workshop on behalf of myself and my co-authors.

Abstract: The FORTRESS (Feature Optimization and Robustness Techniques for 3D Detection Systems) method introduces a novel approach to enhancing the robustness of 3D object detection in autonomous driving. Building on the RayDN architecture, FORTRESS incorporates a modified EVA ViT-Large backbone, pre-trained on ImageNet, to achieve deep and resilient feature extraction. The method is further enhanced with a strategic combination of Augmix and DeepAug data augmentation techniques, carefully crafted to address diverse environmental changes and maintain robustness against real-world data distribution shifts. The training process is systematically structured, progressing from clean datasets to increasingly complex scenarios, each phase contributing to the development of a more robust detection system. By adopting feature optimization and robustness techniques, FORTRESS not only refines the detection capabilities but also ensures the model's adaptability to varied and unforeseen environmental conditions. Preliminary results have demonstrated the method's potential as an effective solution for robust BEV detection challenges in autonomous driving. Additionally, FORTRESS was validated in the ICRA 2024 RoboDrive Challenge, where it achieved second place in Track 1: Robust BEV Detection.

Submission Number: 11