SelfLoc

SelfLoc: High Quality Unsupervised Object Localization with Self-Prompt SAM

---

Installation

Dependencies

This code was implemented with python 3.8, PyTorch 1.13.1 and CUDA 12.2. Please install PyTorch. In order to install the additionnal dependencies, please launch the following command:

# Create conda environment
conda env create -f environment.yml

Please Install SAM. The SAM model can be installed using the following commands:

pip install git+https://github.com/facebookresearch/segment-anything.git

Please also Install DINO. The DINO model can be installed using the following commands:

git clone https://github.com/facebookresearch/dino.git
cd dino; 
touch __init__.py
echo -e "import sys\nfrom os.path import dirname, join\nsys.path.insert(0, join(dirname(__file__), '.'))" >> __init__.py; cd ../;

Model

• The self-prompt generator can be found in self_prompt_generator.py.

• The filtering implementation can be found in filter.py.

• The merge strategy can be found in merge.py.

Evaluation

Unsupervised Saliency Detection

python eval_selfloc.py --dataset_name datasetname --mask_root /path/to/gt/masks --pred_root /path/to/predictions

Unsupervised Single Object Discovery

python eval_selfloc_uod.py --dataset_dir /path/to/dataset --dataset_name datasetname --predict_folder /path/to/predictions --output_dir /path/to/output

Unsupervised Camouflaged Object Segmentation

python eval_selfloc.py --dataset_name datasetname --mask_root /path/to/gt/masks --pred_root /path/to/predictions

Benchmark Results

Table 1: Comparison of our SelfLoc and state-of-the-art unsupervised methods on salient object segmentation benchmark datasets. The best and the second best results of each row are highlighted.

Table 2: Comparison of our SelfLoc and state-of-the-art unsupervised methods on single object discovery benchmark datasets. C20k refers to COCO20K dataset

Table 3: Comparison of our SelfLoc and state-of-the-art unsupervised methods on camouflaged object segmentation benchmark datasets. The best and the second best results of each row are highlighted

Qualitative results of unsupervised object discovery and localization. We superimpose the masks generated by Ours, Found, and TokenCut onto images from DUTS, DUT-OMRON, ECSSD, and VOC07 datasets.