helblazer811 / RefSAM

Referring Image Segmentation Benchmarking with Segment Anything Model (SAM)

RefSAM

This repository is for evaluating the basic performance of SAM on the Referring Image Segmementation task. Check out the SAM project here.

The Naive Zero Shot Approach

The very basic approach we use is to:

1. Produce a referring expression representation using the CLIP language transformer.
2. Extract SAM masks from an image.
3. Embed the masked sections into a CLIP model to produce a representation of the section.
4. Compare the masked section representation to the representation of the referring expression.

The code for the approach can be found in model.py

Setup

Install SAM

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

Load the SAM model in the pretrained/ folder

I used the sam_vit_h_4b8939.pth model from the SAM repository. It can be found here

Load the data

Follow the directions in prepare_dataset.md to download and setup the evaluation dataset.

Run the evaluation

To evaluate the approach run.

    python evaluate_on_refcoco.py