// the find
open-mmlab/mmsegmentation
OpenMMLab Semantic Segmentation Toolbox and Benchmark.
MMSegmentation is OpenMMLab's semantic segmentation toolbox, bundling 40+ methods (FCN through SegFormer, Mask2Former, SAN) behind a shared config system and model zoo. It's for researchers and engineers who want to reproduce published segmentation results or swap backbones/heads without rewriting training loops from scratch.
The config inheritance system (_base_ datasets/models/schedules) lets you compose a new experiment by changing a few lines instead of copy-pasting a training script - you get DeepLabV3+ on Cityscapes vs ADE20K by swapping one import. The model zoo ships pretrained weights and benchmark numbers for nearly every method, so you can verify your setup against the paper's reported mIoU instead of trusting a clean training curve. It's grown past plain RGB segmentation into depth estimation (VPD, Adabins) and open-vocabulary segmentation (SAN, CAT-Seg), so the same framework covers adjacent tasks.
Last pushed August 2024 - over a year with no commits means no 2025 architectures and a growing gap between what's in configs/ and current SOTA. The MMEngine/MMCV/MMPreTrain dependency chain is version-fragile; mismatched MMCV builds are a recurring source of opaque import failures that the FAQ exists mostly to work around. Reading any single training config means tracing 3-4 levels of _base_ inheritance first, which is a real onboarding cost before you can tell what a run actually does. It's a training/benchmarking framework, not a deployment path - getting a trained model into production inference means handing off to the separate MMDeploy repo.