MixCycle: Mixup Assisted Semi-Supervised 3D Single Object Tracking with Cycle Consistency

Qiao WuJiaqi YangKun SunChu'ai ZhangYanning ZhangMathieu Salzmann

Mar 2023

摘要

3D single object tracking (SOT) is an indispensable part of automateddriving. Existing approaches rely heavily on large, densely labeled datasets.However, annotating point clouds is both costly and time-consuming. Inspired bythe great success of cycle tracking in unsupervised 2D SOT, we introduce thefirst semi-supervised approach to 3D SOT. Specifically, we introduce twocycle-consistency strategies for supervision: 1) Self tracking cycles, whichleverage labels to help the model converge better in the early stages oftraining; 2) forward-backward cycles, which strengthen the tracker's robustnessto motion variations and the template noise caused by the template updatestrategy. Furthermore, we propose a data augmentation strategy named SOTMixupto improve the tracker's robustness to point cloud diversity. SOTMixupgenerates training samples by sampling points in two point clouds with a mixingrate and assigns a reasonable loss weight for training according to the mixingrate. The resulting MixCycle approach generalizes to appearance matching-basedtrackers. On the KITTI benchmark, based on the P2B tracker, MixCycle trainedwith $\textbf{10%}$ labels outperforms P2B trained with $\textbf{100%}$ labels,and achieves a $\textbf{28.4%}$ precision improvement when using $\textbf{1%}$labels. Our code will be publicly released.

展开全部

图表提取