AdvSemiSeg

Why we included this project

Labeled data is the bottleneck in most semantic segmentation work, and this repo is a compact demonstration of one way around it: train an adversarial discriminator alongside the segmentation network so unlabeled images get pulled into training. The method comes from the BMVC 2018 paper and works by pushing predictions to look like ground-truth masks, which lets a small pool of labeled images plus a much larger stack of unlabeled ones still produce a usable model. The code is a faithful PyTorch port built on a DeepLabv2-ResNet101 baseline, with pretrained models for several label ratios (1/8, 1/4, 1/2 and fully supervised) and an evaluation script that reports per-class IoU on VOC2012. That makes it easy to reproduce the paper's numbers, or to study the adversarial loop in isolation without pulling in a large framework.