Official implementation of 'TiG-BEV: Multi-view BEV 3D Object Detection via Target Inner-Geometry Learning'.
We propose TiG-BEV, a learning scheme of Target Inner-Geometry from the LiDAR modality into camera-based BEV detectors for both dense depth and BEV features. First, we introduce an inner-depth supervision module to learn the low-level relative depth relations between different foreground pixels. This enables the camerabased detector to better understand the object-wise spatial structures. Second, we design an inner-feature BEV distillation module to imitate the high-level semantics of different keypoints within foreground targets. To further alleviate the BEV feature gap between two modalities, we adopt both inter-channel and inter-keypoint distillation for feature-similarity modeling. With our target inner-geometry distillation, TiG-BEV can effectively boost BEVDepth by +2.3% NDS and +2.4% mAP, along with BEVDet by +9.1% NDS and +10.3% mAP on nuScenes val set.
| Method | mAP | NDS |
|---|---|---|
| TiG-BEV-R50 | 33.8 | 37.5 |
| TiG-BEV4D-R50 | 36.6 | 46.1 |
We provide the model and log of TiG-BEV4D-R101-CBGS.
| Method | mAP | NDS | Model | Log |
|---|---|---|---|---|
| TiG-BEV4D-R101-CBGS | 44.0 | 54.4 |
Please see getting_started.md in BEVDet.
We sincerely thank these great open-sourced work below:
If you find this project useful, please cite:
@article{huang2022tig,
title={TiG-BEV: Multi-view BEV 3D Object Detection via Target Inner-Geometry Learning},
author={Huang, Peixiang and Liu, Li and Zhang, Renrui and Zhang, Song and Xu, Xinli and Wang, Baichao and Liu, Guoyi},
journal={arXiv preprint arXiv:2212.13979},
year={2022}
}