qianduoduolr / FGVC

[ICCV 2023] Learning Fine-Grained Features for Pixel-wise Video Correspondences

[ICCV 2023] Learning Fine-Grained Features for Pixel-wise Video Correspondences

Rui Li¹, Shenglong Zhou¹, and Dong Liu¹

¹University of Science and Technology of China, Hefei, China

[Paper] / [Demo] / [Project page] / [Poster] / [Intro]

This is the official code for "Learning Fine-Grained Features for Pixel-wise Video Correspondences".

Without any fine-tuning, the proposed method can be directly applied to various correspondence-related tasks including long-term point tracking, video object segmentation, etc.

⭐ News

• 2023.07.14: Our paper "Learning Fine-Grained Features for Pixel-wise Video Correspondences" is accepted to ICCV 2023. The code for inference and training will be released as soon as possible.

• 2023.10.04: We have released the code and models of the paper "Learning Fine-Grained Features for Pixel-wise Video Correspondences".

Overview

Video analysis tasks rely heavily on identifying the pixels from different frames that correspond to the same visual target. To tackle this problem, recent studies have advocated feature learning methods that aim to learn distinctive representations to match the pixels, especially in a self-supervised fashion. Unfortunately, these methods have difficulties for tiny or even single-pixel visual targets. Pixel-wise video correspondences were traditionally related to optical flows, which however lead to deterministic correspondences and lack robustness on real-world videos. We address the problem of learning features for establishing pixel-wise correspondences. Motivated by optical flows as well as the self-supervised feature learning, we propose to use not only labeled synthetic videos but also unlabeled real-world videos for learning fine-grained representations in a holistic framework. We adopt an adversarial learning scheme to enhance the generalization ability of the learned features. Moreover, we design a coarse-to-fine framework to pursue high computational efficiency. Our experimental results on a series of correspondence-based tasks demonstrate that the proposed method outperforms state-of-the-art rivals in both accuracy and efficiency.

Citation

If you find this repository useful for your research, please cite our paper:

@inproceedings{li2023learning,
  title={Learning Fine-Grained Features for Pixel-wise Video Correspondences},
  author={Li, Rui and Zhou, Shenglong and Liu, Dong},
  booktitle={ICCV},
  pages={9632--9641},
  year={2023}
}

Our other paper related to video correspondence learning (Spa-then-Temp):

@inproceedings{li2023spatial,
  title={Spatial-then-Temporal Self-Supervised Learning for Video Correspondence},
  author={Li, Rui and Liu, Dong},
  booktitle={CVPR},
  pages={2279--2288},
  year={2023}
}

Prerequisites

• Python 3.8.8
• PyTorch 1.9.1
• mmcv-full == 1.5.2
• davis2017-evaluation

To get started, first please clone the repo

git clone https://github.com/qianduoduolr/FGVC

For convenience, we provide a Dockerfile. Alternatively, you can install all required packages manually. Our code is based on mmcv framework and Spa-then-Temp. You can refer to those repositories for more information.

Data Preparation

Please refer to README.md for more details.

Evaluation

Method	BADJA / PCK@0.2	JHMDB / PCK@0.1	TAP-DAVIS / < D	TAP-Kinetics / < D
RAFT	45.6	66.4	42.1	44.3
TAPNet	-	62.3	48.6	54.4
PIPs	62.3	-	55.3	48.2
Ours	69.7	66.8	62.8	54.6

The evaluation is particularly conducted on pixel-wise correspondence-related tasks, i.e., point tracking, on TAP-Vid dataset, JHMDB, and BADJA. The results are shown above.

We follow the prior studies to leverage label propagation for inference, which can be achieved by:

bash tools/dist_test.sh ${CONFIG}  ${GPUS} ${TASK} ${CKPT}

Note you need download the pre-trained models with this link for the CKPT. Note the TASK consists of 'davis' (for TAP-Vid-DAVIS), 'kinetics' (for TAP-Vid-Kinetics), 'jhmdb' (for human keypoint tracking), and 'badja' (for animal keypoint tracking).

We give a inference cmd example:

# testing for point tracking on TAP-Vid-DAVIS with 4 GPUs
bash tools/dist_test.sh configs/eval/res18_d1_eval.py 4 davis ckpt/res18_d1_fly_ytv_mixed_training.pth

The results will be saved to eval/. Please note we do inference on 4 A100 GPUs, which has 80G memory. Here we give the inference code to support other GPUs with smaller memory size, which may cost more time for inference, we plan to give a more efficient version of the inference code with label propagation later. If you have enough memory, you can simply increase the step of test_cfg in CONFIG for faster inference in current version.

Tranining

We perform training on FlyingThings and YouTube-VOS. Before training you need to download the pre-trained 2D encoder from this link, and modify the pretrained in model.teacher in the config. You can also try more stronger model pre-trained on large-scale image dataset, i.e., MoCo, DetCo, which may get better results.

bash tools/dist_train.sh configs/train/mixed_train_res18_d1_l2_rec_ytv_fly.py 4

License

This work is licensed under MIT license. See the LICENSE for details.

Acknowledgement

The codebase is implemented based on the MMCV, tapnet, pips, and VFS. Thanks for these excellent open source repositories.