Home

Awesome

Event Camera Data Pre-training<br><sub>Official PyTorch Implementation of the ICCV 2023 Paper</sub><br><sub>Paper | Supp | Video </sub><br>

<p align="left"> <img src="doc/teaser.jpg"/> <br> </p>

Getting Started

1. Dependency

2. Datasets

└───N_Imagenet
    └───/extracted_train/n******/***.npz
    │
    └───/extracted_val/n******/***.npz
 ___/train/n******/***.JPEG
│
└───/val/n******/***.JPEG

3. Pre-training

python3 extract_clip_feature.py --batch 128 --source_dir #where you store ImageNet --save_dir #where you store the CLIP features
python3 train_bert.py --opt config/ours/pr_vits.yml --gpus 4 --num_nodes 4

4. Linear probing / Fine-tuing

python3 train_bert.py --opt config/ours/lr_vits.yml --gpus 4 --num_nodes 4
python3 train_bert.py --opt config/ours/ft_vits.yml --gpus 4 --num_nodes 4

Pre-trained models

Find pre-trained models here.

Contact

If you have any questions, please drop me an email.

Acknowledgement

Most of the code is borrowed from:

Citation

If you find this code useful, please consider citing:

@article{Yang_2023,
  author       = {Yan Yang and
                  Liyuan Pan and
                  Liu Liu},
  title        = {Event Camera Data Pre-training},
  journal      = {CoRR},
  volume       = {abs/2301.01928},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2301.01928},
  doi          = {10.48550/arXiv.2301.01928},
  eprinttype    = {arXiv},
  eprint       = {2301.01928},
}