JH-GEECS/visual_token_matching_JH

[ICLR'23 Oral] Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching

Visual Token Matching

(News) Our paper received the Outstanding Paper Award in ICLR 2023!

This repository contains official code for Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching (ICLR 2023 oral).

Setup

1. Download Taskonomy Dataset (tiny split) from the official github page https://github.com/StanfordVL/taskonomy/tree/master/data.

• You may download data of depth_euclidean, depth_zbuffer, edge_occlusion, keypoints2d, keypoints3d, normal, principal_curvature, reshading, segment_semantic, and rgb.
• (Optional) Resize the images and labels into (256, 256) resolution.
• To reduce the I/O bottleneck of dataloader, we stored data from all buildings in a single directory. The directory structure looks like:

<root>
|--<task1>
|   |--<building1>_<file_name1>
|   | ...
|   |--<building2>_<file_name1>
|   |...
|
|--<task2>
|   |--<building1>_<file_name1>
|   | ...
|   |--<building2>_<file_name1>
|   |...
|
|...

2. Create data_paths.yaml file and write the root directory path (<root> in the above structure) by taskonomy: PATH_TO_YOUR_TASKONOMY.

3. Install pre-requirements by pip install -r requirements.txt.

4. Create model/pretrained_checkpoints directory and download BEiT pre-trained checkpoints to the directory.

• We used beit_base_patch16_224_pt22k checkpoint for our experiment.

Usage

Training

python main.py --stage 0 --task_fold [0/1/2/3/4]

Fine-tuning

python main.py --stage 1 --task [segment_semantic/normal/depth_euclidean/depth_zbuffer/edge_texture/edge_occlusion/keypoints2d/keypoints3d/reshading/principal_curvature]

Evaluation

python main.py --stage 2 --task [segment_semantic/normal/depth_euclidean/depth_zbuffer/edge_texture/edge_occlusion/keypoints2d/keypoints3d/reshading/principal_curvature]

After the evaluation, you can print the test results by running python print_results.py

References

Our code refers the following repositores:

• Taskonomy
• timm
• BEiT: BERT Pre-Training of Image Transformers
• Vision Transformers for Dense Prediction
• Inverted Pyramid Multi-task Transformer for Dense Scene Understanding
• Hypercorrelation Squeeze for Few-Shot Segmentation
• Cost Aggregation with 4D Convolutional Swin Transformer for Few-Shot Segmentation

Citation

If you find this work useful, please consider citing:

@inproceedings{kim2023universal,
  title={Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching},
  author={Donggyun Kim and Jinwoo Kim and Seongwoong Cho and Chong Luo and Seunghoon Hong},
  booktitle={International Conference on Learning Representations},
  year={2023},
  url={https://openreview.net/forum?id=88nT0j5jAn}
}

Acknowledgements

The development of this open-sourced code was supported in part by the National Research Foundation of Korea (NRF) (No. 2021R1A4A3032834).