juz031 / selfmask Goto Github PK

      ___           ___                         ___         ___           ___           ___           ___     
     /  /\         /  /\                       /  /\       /__/\         /  /\         /  /\         /__/|    
    /  /:/_       /  /:/_                     /  /:/_     |  |::\       /  /::\       /  /:/_       |  |:|    
   /  /:/ /\     /  /:/ /\    ___     ___    /  /:/ /\    |  |:|:\     /  /:/\:\     /  /:/ /\      |  |:|    
  /  /:/ /::\   /  /:/ /:/_  /__/\   /  /\  /  /:/ /:/  __|__|:|\:\   /  /:/~/::\   /  /:/ /::\   __|  |:|    
 /__/:/ /:/\:\ /__/:/ /:/ /\ \  \:\ /  /:/ /__/:/ /:/  /__/::::| \:\ /__/:/ /:/\:\ /__/:/ /:/\:\ /__/\_|:|____
 \  \:\/:/~/:/ \  \:\/:/ /:/  \  \:\  /:/  \  \:\/:/   \  \:\~~\__\/ \  \:\/:/__\/ \  \:\/:/~/:/ \  \:\/:::::/
  \  \::/ /:/   \  \::/ /:/    \  \:\/:/    \  \::/     \  \:\        \  \::/       \  \::/ /:/   \  \::/~~~~ 
   \__\/ /:/     \  \:\/:/      \  \::/      \  \:\      \  \:\        \  \:\        \__\/ /:/     \  \:\     
     /__/:/       \  \::/        \__\/        \  \:\      \  \:\        \  \:\         /__/:/       \  \:\    
     \__\/         \__\/                       \__\/       \__\/         \__\/         \__\/         \__\/    

This repo contains the code to reproduce the experiments results in the paper "Unsupervised Salient Object Detection with Spectral Cluster Voting". [Project page]

• Demo
• Preparation
• Training
• Inference
• Pre-trained weights
• Generating pseudo-masks with own images
• Citation
• Acknowledgements

Please find our demo built with Hugging Face and Gradio.

To train/evaluate SelfMask, you first need to download some datasets. For training, please download the DUTS-TR dataset and its pseudo-masks (located at datasets/swav_mocov2_dino_p16_k234.json in this repo). For evaluation, please download the DUT-OMRON, DUTS-TE, and ECSSD datasets. Please don't change the (sub)directory name(s) as the code assumes the original directory names. We advise you to put the downloaded dataset(s) into the following directory structure for ease of implementation:

your_dataset_directory
├──DUTS
│  ├──DUTS-TE-Image
│  ├──DUTS-TE-Mask
│  ├──DUTS-TR-Image
├──DUTS-OMRON
│  ├──DUT-OMRON-image
│  ├──pixelwiseGT-new-PNG
├──ECSSD
   ├──images
   ├──ground_truth_mask

faiss-gpu==1.7.1
torch>=1.10
matplotlib==3.5.1
natsort==7.1.1
opencv==4.5.5
pycocotools==2.0.4
scikit-learn==1.0.2
scipy==1.7.3
timm==0.4.12
tqdm==4.63.0
ujson==4.2.0
wandb==0.12.11
pyyaml==6.0

Before running a training script, you need to set up some directory/file paths (e.g., dataset directory). For this please open "duts-dino-k234-nq20-224-swav-mocov2-dino-p16-sr10100.yaml" file in configs directory and find "dir_ckpt", "dir_dataset", and "pseudo_masks_fp" arguments. Then, type your corresponding paths:

...
dir_ckpt: [YOUR_DESIRED_CHECKPOINT_DIR]
...
dir_dataset: [YOU_DATASET_DIR]
...
pseudo_masks_fp: [PATH_TO_DOWNLOADED_PSEUDO_MASKS_FILE]
...

To train a model with 20 queries from scratch, please move to the scripts directory and run

bash train-selfmask-nq20.sh

It is worth noting that, by default, the code will evaluate the model at the end of every epoch, stores the metric values (and the model weights if there was an improvement in terms of metrics).

To run an inference of a pre-trained model, please run

python3 evaluator.py --dataset_name $DATASET_NAME --p_state_dict $PATH_TO_WEIGHTS --config $PATH_TO_MODEL_CONFIG

Here, the config file is the configuration file used for pre-training.

We provide the pre-trained weights used for our experiments:

IoUs are measured on the DUTS-TE benchmark.

To generate pseudo-masks for your own images, please use mask_generator.py file. All you need to do is to set a list of image paths in the script (and save the resulting file if needed).

@InProceedings{shin2022selfmask,
    author    = {Shin, Gyungin and Albanie, Samuel and Xie, Weidi},
    title     = {Unsupervised Salient Object Detection With Spectral Cluster Voting},
    booktitle = {CVPRW},
    year      = {2022}
}

We borrowed the code for ViT, DINO, and MaskFormer from https://github.com/rwightman/pytorch-image-models, https://github.com/facebookresearch/dino, and https://github.com/facebookresearch/MaskFormer, respectively.

If you have any questions, please contact us at gyungin [at] robots [dot] ox [dot] ac [dot] uk.