Stereo Neural Vernier Caliper

We propose a new object-centric framework for learning-based stereo 3D object detection. Previous studies build scene-centric representations that do not consider the significant variation among outdoor instances and thus lack flexibility functionalities an instance-level model can offer. such by formulating tackling local update problem, i.e., how to predict refined given initial cuboid guess. demonstrate solving this problem complement approaches in (i) building coarse-to-fine multi-resolution system, (ii) performing model-agnostic location refinement, (iii) conducting tracking-by-detection. Extensive experiments effectiveness of our approach, which achieves state-of-the-art performance on KITTI benchmark. Code pre-trained models are available at https://github.com/Nicholasli1995/SNVC.