Motion-to-Matching: A Mixed Paradigm for 3D Single Object Tracking

This is the official code release of "Motion-to-Matching: A Mixed Paradigm for 3D Single Object Tracking"

Abstract

3D single object tracking with LiDAR points is an important task in the computer vision field. Previous methods usually adopt the matching-based or motion-centric paradigms to estimate the current target status. However, the former is sensitive to the similar distractors and the sparseness of point clouds due to relying on appearance matching, while the latter usually focuses on short-term motion clues (eg. two frames) and ignores the long-term motion pattern of target. To address these issues, we propose a mixed paradigm with two stages, named MTM-Tracker, which combines motion modeling with feature matching into a single network. Specifically, in the first stage, we exploit the continuous historical boxes as motion prior and propose an encoder-decoder structure to locate target coarsely. Then, in the second stage, we introduce a feature interaction module to extract motion-aware features from consecutive point clouds and match them to refine target movement as well as regress other target states. Extensive experiments validate that our paradigm achieves competitive performance on large-scale datasets (70.9% in KITTI and 51.70% in NuScenes).

Method

Performance

Tip: Open source performance is trained and tested on NVIDIA RTX 4090. If you test on NVIDIA RTX 3090, the overall performance will also be higher than our paper, but slightly lower than 4090.

Setup

Install basic environment

conda create -n mtm python=3.9 -y
conda activate mtm

# please ensure that the pytorch version is correct
pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html

# please refer to https://github.com/traveller59/spconv
pip install spconv-cu113

git clone https://github.com/LeoZhiheng/MTM-Tracker.git
cd MTM-Tracker/
pip install -r requirements.txt

python setup.py develop

Please ensure that the Pytorch version is consistent with ours, otherwise there may be performance differences.

Install other libraries

You can find all third-party libraries in "ltr/ops".

cd ltr/ops/deformattn/
python setup.py develop

cd ltr/ops/points_op/
python setup.py develop

Dataset preparation

Download the dataset from KITTI Tracking and organize the downloaded files as follows:

MTM-Tracker                                           
|-- data                                     
|   |-- kitti                                                                          
│   │   └── training
│   │       ├── calib
│   │       ├── label_02
│   │       └── velodyne

QuickStart

Training

For training, you can customize the training by modifying the parameters in the yaml file of the corresponding model, such as 'CLASS_NAMES'. But we recommend using the default yaml files to ensure optimal performance.

After configuring the yaml file, run the following command to start training.

cd tools/
# for a single GPU
python train_track.py --batch_size 32 --epoch 40 --cfg_file $model_config_path --fix_random_seed --extra_tag base
# for example
python train_track.py --batch_size 32 --epoch 40 --cfg_file cfgs/kitti/mtm-ped.yaml --fix_random_seed --extra_tag base

# for multiple GPUs
python -m torch.distributed.launch --nproc_per_node=2 train_track.py --launcher pytorch --batch_size 32 --epoch 40 --cfg_file $model_config_path --fix_random_seed --sync_bn --extra_tag base

Evaluation

The pre-trained models can be downloaded at this Link. Then, you can create a file called "model" in the root directory and put all models into it.

cd tools/
# for a single model
python test_track.py --cfg_file $model_config_path --ckpt $your_saved_ckpt --extra_tag base
# for example
python test_track.py --cfg_file cfgs/kitti/mtm-ped.yaml --ckpt ../model/ped.pth --extra_tag base

# for all saved models
python test_track.py --cfg_file $model_config_path --ckpt $your_saved_ckpt --extra_tag base --eval_all
# for example
python test_track.py --cfg_file cfgs/kitti/mtm-ped.yaml --ckpt ../output/kitti/mtm-ped --extra_tag base --eval_all

To find the best model, we suggest setting “--eval all” and testing the last 15 checkpoints. (We default to saving the last 15 models during training.)

Acknowledgment

• This repo is built upon OpenPCDet.
• Thank traveller59 for his implementation of Spconv.
• Thank tianweiy for his implementation of CenterPoint.
• Thank 3bobo for his implementation of LTTR.
• Thank jiashunwang for his implementation of MRT.

Citation

If you find the project useful for your research, you may cite,

@article{li2023motion,
  title={Motion-to-Matching: A Mixed Paradigm for 3D Single Object Tracking},
  author={Li, Zhiheng and Lin, Yu and Cui, Yubo and Li, Shuo and Fang, Zheng},
  journal={IEEE Robotics and Automation Letters},
  year={2023},
  publisher={IEEE}
}