CoinCheung/BiSeNet

What it solves

This project provides a high-performance implementation of BiSeNetV1 and BiSeNetV2, which are deep learning architectures designed for real-time semantic segmentation. It addresses the need for accurate image segmentation that can run at high frame rates on various hardware platforms.

How it works

The project implements two versions of the BiSeNet architecture, enabling the model to categorize every pixel in an image into a specific class (e.g., road, car, pedestrian). It includes tools for training, evaluating, and fine-tuning these models on standard datasets like Cityscapes, COCO-Stuff, and ADE20K. To ensure efficiency in production, the project provides deployment guides for TensorRT, NCNN, OpenVINO, and Triton Inference Server.

Who it’s for

This is intended for computer vision engineers and researchers who need to implement real-time semantic segmentation in applications such as autonomous driving or robotics, where low latency and high throughput are critical.

Highlights

• Real-time performance: Achieves high FPS (up to 426 FPS in int8 precision) on the Cityscapes dataset.
• Multi-platform deployment: Includes explicit support for TensorRT, NCNN, OpenVINO, and Triton Inference Server.
• Flexible evaluation: Supports single-scale, multi-scale, and cropped evaluation methods.
• Pre-trained weights: Provides downloadable weights for multiple benchmark datasets.