A segmentation reimplementation that admits its variance

What it does This is a PyTorch implementation of BiSeNetV1 and BiSeNetV2, the bilateral segmentation networks for semantic segmentation. It provides pretrained weights and mIOU benchmarks across Cityscapes, COCO-Stuff, and ADE20K, along with training and finetuning scripts for custom datasets. The project also bundles export and runtime support for TensorRT, NCNN, OpenVINO, and Triton Inference Server.

The interesting bit The README is unusually candid about the rough edges: BiSeNetV2 has fewer FLOPs yet trains slower than V1, repeated runs can swing mIOU by roughly two points on Cityscapes, and some benchmarks are admitted “make it work” efforts where no official settings exist. That transparency is rarer than the architecture.

Key highlights

• Pretrained weights and mIOU/fps tables for Cityscapes, COCO-Stuff, and ADE20K
• Deployment targets include TensorRT, NCNN, OpenVINO, and Triton Inference Server
• Supports inference on images, video, and camera streams
• Refactored codebase with custom dataset support via config files
• Multi-GPU training and evaluation support

Caveats

• FPS and some benchmarks were measured differently from the original papers; COCO-Stuff uses the 123k split rather than the authors’ 10k split
• The author warns of high training variance, with BiSeNetV2 Cityscapes results varying between 73.1 and 75.1 mIOU across runs
• ADE20K numbers are explicitly labeled as unoptimized “make it work” results with no official baseline to compare against

Verdict Grab it if you need a segmentation workhorse that prioritizes deployment and candor over leaderboard chasing. Look elsewhere if you need guaranteed reproducibility down to the decimal point or official dataset splits for strict comparison.