ZhengPeng7/BiRefNet

What it solves

BiRefNet is designed for high-resolution dichotomous image segmentation (DIS), which involves accurately separating a foreground object from its background. It addresses the challenge of maintaining high precision and detail in the segmentation masks, especially for high-resolution images where traditional methods may struggle with boundary accuracy.

How it works

The project implements a bilateral reference mechanism to improve segmentation performance. It provides a variety of pre-trained weights tailored for different tasks, including general-purpose segmentation, matting, and specific benchmarks like DIS, Camouflaged Object Detection (COD), and High-Resolution Salient Object Detection (HRSOD). The model can be loaded via Hugging Face and supports various input resolutions, including dynamic resolution ranges and specialized versions for 2K and 2048x2048 images.

Who it’s for

This tool is intended for researchers and developers working in computer vision, specifically those needing high-quality background removal or object segmentation from high-resolution images.

Highlights

• SOTA Performance: Achieves state-of-the-art results on multiple benchmarks including DIS, COD, and HRSOD.
• Versatile Model Zoo: Offers specialized models for general use, trimap-free matting, and high-resolution (HR) tasks.
• Flexible Resolution Support: Includes a BiRefNet_dynamic version that is robust across resolutions from 256x256 to 2304x2304.
• Efficient Inference: Supports FP16 inference for better efficiency, achieving up to 17 FPS at 1024x1024 resolution on an RTX 4090.
• Easy Integration: Can be loaded in a single line of code using the transformers library.