PP-DocLayout_plus-L ONNX (layout)

Serving repo for the PaddleOCR PP-DocLayout_plus-L page-layout model (RT-DETR-L, 20 classes, mAP 83.2 on the Paddle sub-area set).

Provenance

• pp_doclayout_plus_l.onnx: byte-identical to PaddlePaddle/PP-DocLayout_plus-L_onnx inference.onnx (Apache-2.0), sha256:77afb2caa74dd13240d087d2eced91d7fcd2caebd16006a0a66162fc8707ff0e.
• labels.txt: the 20-class label_list from the upstream inference.yml, verbatim (index = class id).

Serving (bobine layout slot)

Inputs image [1,3,800,800] (direct resize, /255, no ImageNet norm), im_shape [[h,w]], scale_factor [[800/h,800/w]]. Outputs fetch_name_0 [300,6] = [class,score,x0,y0,x1,y1] in original-image coords + fetch_name_1 query count. Operate at score >= 0.5 (upstream draw_threshold).