GF-SAM (Customized Fork)

Note: This is an enhanced version of the original GF-SAM with key improvements for Few-Shot Segmentation.

Key Improvements

1. Targeted Segmentation: Process only specific images instead of the entire dataset.
2. Better Support Images: No duplicate images in support sets.

---

Intallation

Example conda environment setup

conda create --name gfsam python=3.13.2
conda activate gfsam

# my CUDA version is 12.4
pip install torch==2.6.0 torchvision==0.21.0

pip install 'git+https://github.com/facebookresearch/detectron2.git'

git clone https://github.com/raziyev/GF-SAM-custom.git
cd GF-SAM-custom
pip install -r requirements.txt

Model Preparation

• For model weights download, see original GETTING_STARTED.md.

Dataset Preparation

1. Download:

Download datasets using the official guide.

2. Folder Structure:

Organize your files exactly like this:

GF-SAM-custom/          
└── datasets/
    └── FSS-1000/
        ├── data/          # All images/masks go here
        │   ├── scratch/   # Example class folder
        │   │   ├── 1.jpg  # Image file (numeric names only)
        │   │   ├── 1.png  # Mask file (white=object, black=background)
        │   │   ├── 2.jpg 
        │   │   ├── 2.png
        │   │    ...       # more images
        │   ├── dent/      # Another class folder
        │   │   ├── 1.jpg 
        │   │   ├── 1.png
        │   │   ├── 2.jpg
        │   │   ├── 2.png
        │   │   ...
        │   ...
        ├── splits/        # Required for evaluation
        │    └── test.txt  # Controls which folders to process
        │
        └── query_file.txt # Your custom selection file (optional)

3. File Requirements:

• Images:

  • Naming:

    • Only use numbers for filenames (e.g., 1.jpg, 2.png)
    • No letters/symbols in filenames

  • Types:

    • .jpg for input images
    • .png for mask files

  • Mask Colors:

    • Object: Pure white (RGB 255,255,255)
    • Background: Pure black (RGB 0,0,0)

• test.txt configuration:

  • This file determines which class folders to process.

  • Only folders listed in test.txt will be segmented.

  • File location:

  datasets/FSS-1000/splits/test.txt
  

  • File format:

  scratch
  dent
  # Add other folder names here, one per line
  

• query_file.txt configuration:

  • The query_file.txt lets you specify exactly which images to process, rather than segmenting all images in the dataset.

  • File location:

  datasets/FSS-1000/query_file.txt
  

  • File format:

  folder_name  image_number1 image_number2 ...  # (no .jpg extension)
  

  • Example:
    Predicts segmentation masks only for dent/6.jpg, dent/7.jpg, dent/8.jpg and scratch/1.jpg, etc.

  dent 6 7 8
  scratch 1 2 3
  

Usage

How It Works

1. Input Selection:

   • Processes only images specified in query_file.txt
   • Folders must be listed in datasets/FSS-1000/splits/test.txt

2. Support Set Selection:

   • Selects --nshot unique random images from remaining (query image excluded) images in the same folder
   • Ensures no duplicate support images (fixed from original implementation)

3. Output:

   • Implementation log file saved to output/fss/fold0/
   • Visualizations saved to output/vis/ when --visualize 1 is set

Processing Modes

1. Targeted Processing (with query_file.txt):

   • Only processes specified images
   • Example: bus 6 7 8 → processes bus/6.jpg, bus/7.jpg, bus/8.jpg as query

2. Full Dataset Processing (no query file):

   • Processes all images in folders listed in test.txt
   • Uses all images as queries (original behavior)

Command Options

1. Custom Query File

python main_eval.py \
    --datapath datasets \ 
    --benchmark fss \          
    --queryfile my_query.txt \  # Your custom selection (default= None)
    --nshot 1 \                # 1 support image per query
    --fold 0 \
    --log-root "output/fss/fold0" \
    --visualize 1              # Saves visualizations

2. Process All Images

python main_eval.py \
    --datapath datasets \
    --benchmark fss \
    --nshot 1 \
    --fold 0 \
    --log-root "output/fss/fold0"