[layout detection] Task definitions #2009
Description
🚀 The feature
- datasets: DocLayNet, PubLayNet, DocBank, M6Doc, RanLayNet, PRImA, ? (offline work - merge into unique)
- model: https://github.com/Zeba-Xie/RTMDet-R2 (two stages: backbone, neck + head + losses)
- metrics: mmAP (75, 50) - rotated (ref.: https://github.com/open-mmlab/mmrotate/blob/main/mmrotate/core/evaluation/eval_map.py)
- Implement train / eval / latency scripts - reuse DetectionDataset for KIE annotations ?
- Integrate into pipeline - standalone predictor - Extend DocumentBuilder components
- If layout information available improve sorting to keep reading order (heuristic ? graph based ordering (with networkx) ?)
- Needs other ticket: Allow different output formats (markdown, ..)
Motivation, pitch
TODO: Split into single issues - and add better descriptions
Alternatives
No response
Additional context
No response