WI3D: Weakly Incremental 3D Detection via Visual Prompts

Mingsheng Li, Sijin Chen, Shengji Tang, Hongyuan Zhu, Fukun Yin, Tao Chen

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Class-Incremental 3D Object Detection; Continual Learning; Cross-Modal

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Class-incremental 3D object detection demands a 3D detector to locate and recognize novel categories in a stream fashion, while not forgetting its previously learned knowledge. However, existing methods require delicate 3D annotations for learning novel categories, resulting in significant labeling cost. To this end, we explore a label-efficient approach called Weakly Incremental 3D object Detection (WI3D), which teaches a 3D detector to learn new object classes using cost-effective 2D visual prompts. For that, we propose a framework that infuses (i) class-agnostic pseudo label refinement module for high-quality 3D pseudo labels generation, (ii) cross-modal knowledge transfer for representation learning of novel classes, and (iii) reweighting knowledge distillation for preserving old class information. Extensive experiments under different incremental settings on both SUN-RGBD and ScanNet show that our approach learns well to detect novel classes while effectively preserving knowledge of base classes, and surpasses baseline approaches in WI3D scenarios.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2466