README.md

SimpleMatch: A Simple and Strong Baseline for Semantic Correspondence

This is the official code for SimpleMatch implemented with PyTorch.

Prepare Datasets

1. Download PF-PASCAL and PF-WILLOW datasets: Download
2. Download SPair-71k dataset: Download
3. Unzip the datasets and place them in the data directory
4. Organized as follows

./data
    PF-PASCAL
    PF-WILLOW
    SPair-71k

Download pretrained parameters of backbone.

1. DINOv2: Download
2. ibot: Download
3. resnet101: Download

mkdir checkpoints and put them in checkpoints directory.

Environment Settings

conda create -n DCM python=3.8.0
conda activate DCM
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113

pip install -r requirements.txt

Download Pretrained weights

Google Drive

mkdir ckpts

# Then, place the downloaded checkpoints in the ckpts directory

Evaluation on SPair-71k

(To evaluate on other datasets, replace the configuration file and the corresponding checkpoint path.)

python test.py \
    --config configs/task_dinov2-b14_448x448_spair.py \
    --model_path ckpts/dinov2_448x448_spair/best_model.pth  \
    --log_name infer  

Evaluation on AP10k

python test.py --config configs/task_dinov2-b14_448x448_ap10k.py \
    --model_path ckpts/dinvov2_448x448_ap10k/best_model.pth  \
    --log_name infer \
    --cfg-options \
    test_dataloader.dataset.eval_type=cross-family  # `intra-species`, `cross-species`, `cross-family`

Training on SPair-71k

python train.py  \
    --config configs/task_dinov2-b14_252x252_spair.py \
    --work-dir work_dirs/tmp 

Training on PF-PASCAL

python train.py  \
    --config configs/task_ibot-b16_256x256_pfpascal.py \
    --work-dir work_dirs/tmp