This project is not a fork, but it refers to https://github.com/facebookresearch/segment-anything.

Semanta is a Next.js and ONNX template for developers looking to integrate Meta's SAM model for semantic segmentation into their projects with the benefits of ONNX runtime optimizations. It includes basic scaffolding for downloading images, flexibly adding labels, and generating customizable image masks.

• Node.js (our version: v18.14.0)
• ONNX Runtime

git clone https://github.com/cantoramann/semanta.git

npm i

npx next dev

After trying the demo, you may want to override small steps.

To override custom images, refer to the segmentationImagesService.images.setImages() function in app/page.tsx. You load images from buckets such as S3 using the URLs.

The applicaiton requires SAM embeddings (or any other ONNX comptabile image embedding) for each image. This can be achieved by making an external call to your model or containerizing a side service (or a Jupyter Notebook) that interacts with the app. Currently, all embedding URLs are set in the beginning of the project. This may not be efficient and possible for all use cases. This can be flexibly overridden before the runInference() function call in app/page.tsx.

By default, downloaded mask colors are the same as the label colors. To keep all downloaded mask colors in black, you can use the function convertToBlackMask() in components/Segmentor/mask-utils.ts.

To generate embedding, there are multiple models that can be used. The most popular one is SAM. In this resource, Meta shows how to use SAM for image embeddings. Semanta will work with any ONNX compatible embedding.