DocLayNet
Unknown
IBM Research's large-scale document layout analysis dataset with 80,863 annotated pages across 6 document categories: financial reports, scientific papers, patents, government tenders, manuals, and laws & regulations. 11 semantic region labels. The standard benchmark for general-domain document layout segmentation.
Benchmark Stats
SOTA History
Only 4 models on this benchmark
Help build the community leaderboard — submit your model results.
mAP
mAP
Higher is better
| Rank | Model | Source | Score | Year | Paper |
|---|---|---|---|---|---|
| 1 | DocFormerv2-Large DocFormerv2-Large on DocLayNet. 84.1 mAP (COCO-style). Table 2 in DocFormerv2 paper (arXiv 2306.01733, 2023). Adobe Research. | Community | 84.1 | 2026 | Source |
| 2 | DiT-L (Cascade R-CNN) DiT-L with Cascade R-CNN on DocLayNet. 82.6 mAP (COCO-style). Table 4 of DocLayNet paper (arXiv 2206.01062). Best result in the original IBM benchmark paper at publication. | Community | 82.6 | 2026 | Source |
| 3 | LayoutLMv3-Large LayoutLMv3-Large on DocLayNet object detection. 79.5 mAP. Table 3 in LayoutLMv3 paper (arXiv 2204.08387, ACM MM 2022). Microsoft Research. | Community | 79.5 | 2026 | Source |
| 4 | DINO (ResNet-50) DINO detector with ResNet-50 backbone on DocLayNet. 73.4 mAP. Reported in DocLayNet follow-up comparisons as strong detection baseline. arXiv 2203.03605. | Community | 73.4 | 2026 | Source |