Polygon-RNN++: When drawing boundaries is too tedious for humans alone

What it does

Polygon-RNN++ generates polygon outlines around objects in image crops. Feed it a crop and a starting vertex, and it predicts the sequence of points tracing the object’s boundary. The repo contains only inference code—run the Jupyter notebook or shell script, get polygons in the output/ folder. Models were trained on Cityscapes and weigh 448 MB.

The interesting bit

The real product isn’t full automation; it’s human-in-the-loop speedup. The model predicts most vertices, humans correct the mistakes, and the loop tightens. The authors also published a separate PyTorch repo with training code and annotation tools—this TensorFlow release is essentially the frozen demo.

Key highlights

• Pre-trained models available via shell script; CPU runs “albeit slowly”
• Interactive demo and video walkthrough linked from the README
• PyTorch reimplementation with full training/tooling exists elsewhere
• Citation includes both the 2018++ paper and the original 2017 Polygon-RNN
• Jupyter notebook provides step-by-step inference walkthrough

Caveats

• Only inference code here; training requires the PyTorch sibling repo
• Clone instructions still point to davidjesusacu/polyrnn, a fork/redirect, rather than the canonical fidler-lab URL

Verdict

Worth a spin if you’re building annotation pipelines or researching interactive segmentation. Skip it if you need end-to-end training—head to the PyTorch repo instead.