Skip to content

The official repo for [AAAI 2024] "SimDistill: Simulated Multi-modal Distillation for BEV 3D Object Detection""

License

Notifications You must be signed in to change notification settings

ViTAE-Transformer/SimDistill

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SimDistill: Simulated Multi-modal Distillation for BEV 3D Object Detection

Haimei Zhao, Qiming Zhang, Shanshan Zhao, Zhe Chen, Jing Zhang, and Dacheng Tao

News | Abstract | Method | Results | Preparation | Code | Statement

News

  • (2023/3/29) SimDistill is accepted by AAAI 2024!.
  • (2023/3/29) BEVSimDet is released on arXiv.

Other applications of ViTAE Transformer include: image classification | object detection | semantic segmentation | pose estimation | remote sensingimage matting | scene text spotting

Abstract

Multi-view camera-based 3D object detection has become popular due to its low cost, but accurately inferring 3D geometry solely from camera data remains challenging and may lead to inferior performance. Although distilling precise 3D geometry knowledge from LiDAR data could help tackle this challenge, the benefits of LiDAR information could be greatly hindered by the significant modality gap between different sensory modalities. To address this issue, we propose a \textbf{Si}mulated \textbf{m}ulti-modal \textbf{Distill}ation (\textbf{SimDistill}) method by carefully crafting the model architecture and distillation strategy. Specifically, we devise multi-modal architectures for both teacher and student models, including a LiDAR-camera fusion-based teacher and a simulated fusion-based student. Owing to the ``identical'' architecture design, the student can mimic the teacher to generate multi-modal features with merely multi-view images as input, where a geometry compensation module is introduced to bridge the modality gap. Furthermore, we propose a comprehensive multi-modal distillation scheme that supports intra-modal, cross-modal, and multi-modal fusion distillation simultaneously in the Bird's-eye-view space. Incorporating them together, our SimDistill can learn better feature representations for 3D object detection while maintaining a cost-effective camera-only deployment. Extensive experiments validate the effectiveness and superiority of SimDistill over state-of-the-art methods, achieving an improvement of 4.8% mAP and 4.1% NDS over the baseline detector.

Method

the framework figure

Results

Quantitative results on Nuscenes validation set

quantitative figure

Qualitative results

qualitative figure qualitative figure qualitative figure

Preparation

Prerequisites

The code is built with following libraries:

After installing these dependencies, please run this command to install the codebase:

python setup.py develop

Data Preparation

nuScenes

Please follow the instructions from here to download and preprocess the nuScenes dataset. Please remember to download both detection dataset and the map extension (for BEV map segmentation). After data preparation, you will be able to see the following directory structure (as is indicated in mmdetection3d):

mmdetection3d
├── mmdet3d
├── tools
├── configs
├── data
│   ├── nuscenes
│   │   ├── maps
│   │   ├── samples
│   │   ├── sweeps
│   │   ├── v1.0-test
|   |   ├── v1.0-trainval
│   │   ├── nuscenes_database
│   │   ├── nuscenes_infos_train.pkl
│   │   ├── nuscenes_infos_val.pkl
│   │   ├── nuscenes_infos_test.pkl
│   │   ├── nuscenes_dbinfos_train.pkl

Code

Setup

python setup.py develop

Training

different loss items should be changed in configs/nuscenes/det/centerhead/lssfpn/camera/256x704/swint/convfuser.yaml

and different backbone networks can be choosed, including swinT, vitaev2, and bevformer in configs/nuscenes/det/centerhead/lssfpn/camera/256x704/

torchpack dist-run -np 8 python tools/train.py configs/nuscenes/det/centerhead/lssfpn/camera/256x704/swint/convfuser.yaml --data.samples_per_gpu 3 --max_epochs 20 --data.workers_per_gpu 6 --run-dir swinT-twobranchesloss --load_from ../bevfusion-main/pretrained/bevfusion-det.pth

Evaluation

torchpack dist-run -np 8 python tools/test.py configs/nuscenes/det/centerhead/lssfpn/camera/256x704/swint/convfuser.yaml --xxx.pth --eval bbox

Statement

@inproceedings{zhao2024simdistill, title={SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection}, author={Zhao, Haimei and Zhang, Qiming and Zhao, Shanshan and Chen, Zhe and Zhang, Jing and Tao, Dacheng}, booktitle={Proceedings of the AAAI Conference on Artificial Intelligence}, volume={38}, number={7}, pages={7460--7468}, year={2024} }

About

The official repo for [AAAI 2024] "SimDistill: Simulated Multi-modal Distillation for BEV 3D Object Detection""

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published