The study introduces LabelFormer, an efficient and effective approach for refining object trajectory in offboard perception models using LiDAR point clouds. This technology is a promising alternative to human annotations, generating automatic labels from raw data at a fraction of the cost. LabelFormer outperforms existing models by significantly improving accuracy through advanced temporal reasoning capabilities. It separately encodes each frame’s observations, uses self-attention to reason about the trajectory with full temporal context, and finally decodes the refined object size and per-frame poses. The use of LabelFormer for data augmentation has shown improved downstream detection performance.
Publication date: 2 Nov 2023
Project Page: https://waabi.ai/labelformer/
Paper: https://arxiv.org/pdf/2311.01444
