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, across various tasks like salient object detection (SOD) and camouflaged object detection (COD).

How it works

The project implements a bilateral reference mechanism to achieve high-resolution segmentation. It utilizes various backbones, such as Swin Transformer, and provides specialized models for different resolutions and tasks. The system supports FP16 inference for efficiency and can be integrated via Hugging Face for one-line loading. It also supports ONNX conversion for deployment.

Who it’s for

Computer vision researchers and developers who need high-precision background removal or object segmentation for high-resolution images, as well as those working on matting and salient object detection.

Highlights

• SOTA Performance: Achieves state-of-the-art results on multiple benchmarks including DIS, COD, and HRSOD.
• High Resolution Support: Offers specialized models trained for resolutions up to 2K (2560x1440) and dynamic resolution ranges (256x256 to 2304x2304).
• Efficient Inference: Supports FP16 inference, achieving 17 FPS at 1024x1024 resolution on an RTX 4090.
• Versatile Model Zoo: Provides a wide range of pre-trained weights for general use, matting, and specific academic benchmarks.
• Easy Integration: One-line loading via Hugging Face AutoModelForImageSegmentation.