What it does

This is the reference PyTorch code for a BMVC 2018 paper on semi-supervised semantic segmentation. It trains a DeepLabv2-ResNet101 segmenter on partial Pascal VOC2012 labels, using a discriminator network to push the segmenter toward more realistic outputs on unlabeled images. The authors provide pretrained models for 1/8, 1/4, and 1/2 labeled data fractions, plus a fully supervised baseline.

The interesting bit

The trick is the adversarial hinge: the discriminator learns to distinguish real ground-truth segmentations from the model’s predictions, and the segmenter gets penalized for fooling it. This creates a pressure toward plausible label maps even where no labels exist. The README notes the code is “heavily borrowed” from an existing DeepLab implementation — this is research scaffolding, not a from-scratch rebuild.

Key highlights

• Baseline (full supervision): 73.6% mIOU on VOC2012 val; semi-supervised with 1/8 labels: ~68.8%
• Pretrained models auto-download via evaluation script (semi0.125, semi0.25, semi0.5, advFull)
• Requires augmented VOC labels (SegmentationClassAug) for training
• Hyperparameters map directly to Table 5 in the paper
• Evaluation script saves both colorized outputs and per-class IOU breakdown

Caveats

• PyTorch version is pinned to the 0.2–0.4 era; the README promises a 0.4 migration that apparently never landed
• OpenCV 3.3 specifically causes GPU memory leaks in the multithreaded dataloader
• No training data augmentation or CRF post-processing, so numbers sit below contemporary fully-supervised SOTA

Verdict

Worth a look if you’re reproducing 2018 semi-supervised segmentation baselines or studying how GAN-style losses behave on structured prediction. Skip it if you need production-ready training code or modern PyTorch compatibility.