birefnet

BiRefNet Examples

This directory contains examples for using BiRefNet (Bimodal Referencing Network) for various image segmentation tasks including portrait matting, camouflaged object detection, and general segmentation.

Overview

BiRefNet is an advanced image segmentation model that achieves high-precision segmentation through bimodal referencing mechanisms. The model excels in various segmentation tasks including salient object detection, camouflaged object detection, and portrait matting.

Available Model Variants

General Purpose Models

• general: General-purpose model for versatile segmentation tasks across different domains
• general_bb_swin_v1_tiny: General-purpose model with Swin-V1-Tiny backbone for balanced performance and efficiency (default)
• hr_general: High-resolution general segmentation model for processing larger images with fine details
• lite_general_2k: Lightweight general model optimized for 2K resolution images with efficient inference

Specialized Models

• cod: Camouflaged Object Detection (COD) model for segmenting objects that blend with their surroundings
• dis: Dichotomous Image Segmentation (DIS) model for basic foreground/background separation
• hrsod_dhu: High-Resolution Salient Object Detection (HRSOD) model trained on DHU dataset
• massive: Massive model trained on multiple datasets including DIS5K and TE datasets for robust performance

Portrait & Matting Models

• portrait: Specialized portrait segmentation model for high-quality portrait background removal
• matting: Portrait matting model for precise hair and fine detail preservation in portraits
• hr_matting: High-resolution portrait matting model for detailed portrait segmentation on larger images

Quick Start

Input/Output

• --source <SOURCE>: Input source - image path, folder, or video (default: images/liuyifei.png)

Examples

hr_matting (fp16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant hr_matting --dtype fp16

general_bb_swin_v1_tiny (q4f16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant general_bb_swin_v1_tiny --dtype q4f16

cod (fp16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant cod --dtype fp16

dis (q4f16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant dis --dtype q4f16

hrsod_dhu (fp16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant hrsod_dhu --dtype fp16

massive (q4f16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant massive --dtype q4f16

general (fp16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant general --dtype fp16

portrait (q4f16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant portrait --dtype q4f16

matting (fp16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --variant matting --dtype fp16

Custom model file (fp16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --model ../biRefNet-qnt/COD-epoch-125-fp16.onnx --dtype fp16

Custom model file (q4f16)

cargo run -r -F cuda-full --example birefnet -- --device cuda:2 --processor-device cuda:2 --model ../biRefNet-qnt/general-bb-swin-v1-tiny-epoch-232-q4f16.onnx --dtype q4f16