A Keras buffet of segmentation models, from UNet to PSPNet

What it does

keras-segmentation collects FCN, UNet, PSPNet, SegNet and their variants—backed by VGG, ResNet-50 and MobileNet—under one Keras API. It exposes both a Python module and a command-line interface for training, inference, evaluation and webcam video segmentation, plus a few pre-trained PSPNet checkpoints for ADE20K, Cityscapes and Pascal VOC 2012.

The interesting bit

The library goes beyond a simple model zoo with weight-transfer helpers for fine-tuning and a built-in knowledge-distillation pipeline to compress a heavy teacher into a smaller student. You can also inject custom imgaug augmentation pipelines and train on non-RGB data by adjusting the input channel count, which turns an otherwise plain collection of architectures into a modest training toolkit.

Key highlights

• ~20 model variants spanning FCN, UNet, PSPNet and SegNet with multiple encoder backbones.
• Pre-trained PSPNet weights for ADE20K, Cityscapes and Pascal VOC 2012.
• Fine-tuning via weight transfer and knowledge-distillation utilities for model compression.
• CLI covers training, prediction, video inference, evaluation and dataset verification.
• Hooks for custom augmentation and arbitrary input-channel counts.

Caveats

• The README recommends Keras 2.4.3 and TensorFlow 2.4.1, so compatibility with newer releases is unclear.
• Annotation masks must encode class labels in the blue channel, a format quirk that may require pre-processing.
• The provided README is truncated, leaving some features partially undocumented.

Verdict

A solid starting point if you want a quick Keras baseline for semantic segmentation and can tolerate legacy framework versions. Look elsewhere if you need PyTorch or modern transformer-based segmentation.