Glass Segmentation with Fusion of Learned and General Visual Features (L+GNet)

Official implementation of the L+GNet model, proposed in paper Glass Segmentation with Fusion of Learned and General Visual Features.

Installation

Install dependencies with pip

pip install -r requirements.txt

NOTE: to run the models, you will need Hugging Face setup on your system, as well as access to the DINOv3 model repositories in Hugging Face.

Training

Training can be performed via

python dev.py --mode train --dataset-path <path to dataset>

Dataset folder is expected to contain the following structure

dataset
└── train/
    ├── image/
    └── mask/

where the folders contain the images and masks with corresponding filenames.

See the argument parser for full list of flags that can be provided.

Testing

Testing can be performed via:

python dev.py --mode test --dataset-path <path to dataset> --weights-path <path to stored weights> --metrics

Dataset folder is expected to contain the following structure

dataset
└── test/
    ├── image/
    └── mask/

where the folders contain the images and masks with corresponding filenames.

See the argument parser for full list of flags that can be provided.

Inference demo

A minimal example on how to run inference on a single image has been provided in infer.py.

To run the inference with L+GNet:

python infer.py --image-path <path to an image> --weights-path <path to stored weights> --output-folder <path to output folder>

See argument parser for details if you want to specify the DINOv3 backbone size.

Pretrained weights

Weights for L+GNet trained on different datasets (GDD, GSD, Trans10k-stuff, HSO, all four training sets combined) are available behind the following link: weights folder

Prefix in the filename defines the dataset on which the model has been trained on.

Acknowledgement

Some of the code has been created based on the repo DINOv3-Mask2former by GitHub user Carti-97.