A PyTorch zoo for classic semantic segmentation nets

What it does

pytorch-semseg gathers seven influential semantic segmentation architectures—FCN, U-Net, SegNet, PSPNet, ICNet, Link-Net, and FRRN—under one roof, alongside dataloaders for workhorse datasets like Cityscapes, Pascal VOC, and ADE20K. Instead of scattered one-off scripts, you get a single YAML-driven harness: one file defines the model, data pipeline, augmentations, optimizer, and learning-rate schedule. Training, validation, and testing are handled by separate entry points, with optional DenseCRF post-processing thrown in for good measure.

The interesting bit

The real product here is normalization, not novelty. The author absorbed the drudgery of reconciling different paper implementations into a single interchangeable framework, complete with data augmentation, multiple loss functions, and checkpoint resumption. Think of it as a standardized test bench for mid-2010s segmentation research.

Key highlights

• Seven architectures: FCN (32s/16s/8s), U-Net, SegNet, PSPNet, ICNet, Link-Net, and FRRN (A/B).
• Seven dataset loaders: Cityscapes, Pascal VOC, ADE20K, CamVid, MIT Scene Parsing, NYUDv2, and Sun-RGBD.
• YAML config swaps models, optimizers (SGD, Adam, etc.), losses (cross-entropy, bootstrapped, multi-scale), and augmentation pipelines without code changes.
• Supports resumable training, evaluation with flipped-image augmentation, and DenseCRF post-processing at inference time.
• Packaged on PyPI (v0.1.2) and citable via DOI, suggesting it was built with academic reproducibility in mind.

Caveats

• The requirements lock to PyTorch ≥0.4.0 and torchvision 0.2.0, so modern environments will likely complain; this is very much a legacy codebase.
• No accuracy tables, benchmark numbers, or pretrained weight performance stats appear in the README, so baseline expectations are unclear.
• E-Net and RefineNet are listed under “Upcoming” but are absent from the implemented networks, suggesting the catalog may be frozen.

Verdict

Use this if you need a clean, unified reference implementation of classic segmentation nets for teaching, paper reproduction, or archaeology on legacy PyTorch. Avoid it if you want state-of-the-art models, modern framework support, or verified pretrained weights ready for production.