ArtExtract
ArtExtract is a CRNN (Convolutional-Recurrent Neural Network) trained on 81,444 paintings across 29 artistic styles from the WikiArt dataset. Built as a baseline for the HumanAI ArtExtract GSoC 2026 task — the foundation toward detecting hidden paintings and anomalies in art. Style isn’t a local feature, so unlike a plain CNN, the BiGRU layer captures left-to-right and right-to-left relationships across the full canvas width — exactly what style recognition needs.
Key Technologies
- Backbone: EfficientNet-B3 + 2-layer Bidirectional GRU
- Language: Python
- Framework: PyTorch
- Concepts Applied: CRNN architecture, outlier detection, compound scaling, multi-task learning