This repository demonstrates the use of SAM2 (Segment Anything Model 2) to automatically generate object masks for images. SAM2 efficiently processes prompts to generate masks by sampling over the entire image and predicting multiple masks from single-point input prompts.
The class SAM2AutomaticMaskGenerator implements the automatic mask generation capability by:
- Sampling single-point input prompts in a grid over the image.
- Predicting multiple masks from each point.
- Filtering masks using quality metrics and deduplication techniques like non-maximal suppression.
- Offering additional options to improve mask quality through multiple image crops and postprocessing to remove small disconnected regions and holes.
This notebook can run both locally and in Google Colab, with optimizations for GPU use.
To run the SAM2 model locally, first clone the repository and install the required dependencies:
git clone https://github.com/facebookresearch/segment-anything-2.git
cd segment-anything-2
pip install -r requirements.txtOnce the environment is set up, you can use the SAM2AutomaticMaskGenerator class to automatically generate masks for any given image. The generated masks will be filtered for quality and deduplicated for efficient processing.
The SAM2 model can use pretrained weights for generating object masks. Download the model checkpoint from the link below and place it in the appropriate directory:
This project is licensed under the MIT License.