Loop Mining Large-Scale Unlabeled Data for Corner Case Detection in Autonomous Driving

Jiawei Zhao, Yiting Duan, Jinming Su, yangwangwang, Tingyi Guo, Xingyue Chen, Junfeng Luo

Keywords: corner case, pseudo-supervised learning, close-set detection, open-set detection, loop-optimization

TL;DR: In this paper, we introduce an effective pipeline for unsupervised and supervised corner case detection.

Subject: One/few/zero-shot learning for autonomous perception

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: For obstacle detection in road scenes, it is very challenging to detect novel objects that are not seen or barely seen during training. To address this issue, we propose an efficient pipeline for obstacle detection in road scenes based on large-scale unlabeled data. Specifically, we use large-scale unlabeled data to train a closed-set model and a open-set model separately in a pseudo-supervised learning manner, and then iteratively improve the performance of both models through the proposed loop-optimization strategy, which employs some useful tricks to remove false positive detections about corner cases. Experimental evidence demonstrates that our approach achieves new state-of-the-art on the popular CODA dataset.

Submission Number: 8