How our neural networks find polygons to understand your documents better
summary
Pipeline targets 4‑point ID crops and 6‑point passport contours and evaluates methods by critical error rate and IoU. A lightweight segmentation network predicting heat maps with integral/soft‑argmax halves errors versus homography-parameter regression; adding a border mask helps but errors correlate. Tried differentiable rasterization to optimize IoU directly for passports, spent months and rolled it back due to collapsed, self-intersecting polygons. Introducing sampling-argmax (Gumbel sampling) cuts critical errors by about 30%, enabling passport cropping in product.
AI-written summary · about
chapters (12)
- 0:00Neural polygon detection intro
- 1:11Document processing pipeline
- 3:20Why polygons matter
- 4:02Projective estimation limits
- 7:34Heat map key points
- 9:18Integral post improvements
- 11:10Differentiable rasterization
- 14:44Polygon vectorization challenges
- 16:54Fixing multi-peak errors
- 18:02Sampling argmax trick
- 22:43Key takeaways and results
- 23:47Q&A
related talks
49mGraph Based Processing of Big Images
33mWhat and where? - Machine learning for geospatial image analysis
17mGeoAI: a feature extraction tool set using Tensor Flow and Keras...
15mTowards automated mapping of pedestrian path networks at scale
41mLarge Scale Field Delineation
24mTiling & Parallel Processing of Large Images
46mML Use Cases at Think Big Analytics
21mLand Boundary Delineation using U-Nets