Welcome to SeqTrack3d! 🚀 This repository houses the magic behind our simple yet powerful multi-frame point cloud target tracker that leverages an attention mechanism to enhance tracking robustness.
Before we get started, let's set up our environment to ensure everything runs smoothly.
Our project runs on PyTorch, and we recommend using version 2.0 or above. Follow the official guide to install PyTorch.
If you're a fan of Docker, you can streamline the setup by using this 🐋 Docker image: pytorch/pytorch:2.0.1, which saves you the trouble of manually installing PyTorch.
To install the rest of the dependencies, simply clone our repo and run the installation command:
git clone https://github.com/aron-lin/seqtrack3d.git
cd seqtrack3d
pip install -r requirements.txt
Ready to feed your model with data? We utilize the nuscenes and waymo datasets. Follow the setup instructions on Open3DSOT's GitHub to get started.
Getting your SeqTrack3D model up and training is straightforward. Follow these steps to initiate training with your desired configuration. Open your terminal or command line interface and navigate to the project directory. Then, execute the following command:
python main.py --cfg cfgs/seqtrack3d_nuscenes.yaml --batch_size 4 --epoch 20 --seed 42 --tag "Enter_your_custom_tag_here_for_this_training_session"Jumpstart your projects with our pretrained models available in the pretrainedmodel folder. It's all set for integrating our model into your applications!
To test your SeqTrack3D model with a pre-trained checkpoint, simply run the following command in your terminal or command line interface. Ensure you're in the project directory before executing the command:
python main.py --cfg cfgs/seqtrack3d_nuscenes.yaml --checkpoint pretrainedmodel/seqtrack_nuscenes_car_succ_62_prec_71.ckpt --testCheck out the output folder in the root directory for training logs and testing results. Each experiment is neatly organized by the training/testing start time, dataset, and tag.
Special thanks to:
- Open3DSOT for their fantastic tracking framework.
- The team behind attention-is-all-you-need-pytorch for their clear attention mechanism implementation.
- DETR by Facebook Research for the influential ideas that have deeply benefited our project.
Each of these contributions has been pivotal in shaping our work. We're incredibly grateful for the community's shared knowledge and innovation.
Our project is open-sourced under the MIT license. Feel free to explore, modify, and share your innovations with the world.