Awesome

SAM for SOS

This is the Segment Anything Model (SAM) version for our Segment Object System (SOS).

Installation

Please follow the general installation instructions of SAM as described in the original README.md. Additionally, install the following packages to your Python environment:

pip install scikit-image tqdm

Usage

This version of SAM contains two new scripts for applying SAM to images based on the prompts provided by the object priors evaluated in our SOS paper. See also our main SOS git for example prompt files from various object priors.

To apply SAM as part of SOS with object priors other than the Grid, use the applySAM.py script with the following parameters: the path to input images, e.g. COCO's train2017 folder, the path to the SAM checkpoint, the path to the prompt file, and the path for the output directory:

python applySAM.py /data/SOS/coco/train2017 /data/SOS/SAM_checkpoints/sam_vit_h_4b8939.pth /data/SOS/prompts/prompts_DINO.json /data/SOS/segments/segments_DINO.json

For the Grid obejct prior, which is essentially SAM's everything mode, use the applySAM_grid.py script with the same parameters, except for the prompt file, which is not necessary here:

python applySAM_grid.py /data/SOS/coco/train2017 /data/SOS/SAM_checkpoints/sam_vit_h_4b8939.pth /data/SOS/segments/segments_Grid.json