docTR: Layout model training overview

[felixT2K]

base architecture:

• LWDETR (small ~15M params) - modified to learn also rotated representations (sin + cos head output)

 "class_names": [
            "Caption",
            "Footnote",
            "Formula",
            "List-item",
            "Page-footer",
            "Page-header",
            "Picture",
            "Section-header",
            "Table",
            "Text",
            "Title",
        ],

pretrain: 20 epochs
data: ~600k samples
res:
AP@[.5]:
76.68 %
AP@[.75]:
67.76 %
mAP@[.5:.95]:
60.36 %

fine-tune: 15 epochs
data: ~60k samples
res:
AP@[.5]:
84.29 %
AP@[.75]:
76.82 %
mAP@[.5:.95]:
70.49 %

fine-tune: 3 epochs
data: ~60k samples (straight)
res:
AP@[.5]:
84.50 %
AP@[.75]:
77.40 %
mAP@[.5:.95]:
70.83 %