bu44er/PP-DocLayoutV3-FP16-ONNX

PP-DocLayoutV3 FP16 (ONNX)

An unmodified mirror of the FP16 ONNX export of PP-DocLayoutV3, used as the pinned layout model of a local PDF reader. Not affiliated with or endorsed by PaddlePaddle.

Provenance

Upstream weights	PaddlePaddle/PP-DocLayoutV3_safetensors (Apache-2.0)
FP16 export	stefanj0/PP-DocLayoutV3-FP16-ONNX at revision ca5155572513b7832d297255a1f5cab0ff4b0202
File	PP-DocLayoutV3-fp16.onnx, 67,372,587 bytes
SHA256	8bb693ed3b5dcc1cf926b15d89dfe6abf62bc11cdd0afd33c8ffe039db6f8209

The file is byte-identical to the source export. The export runs the backbone, encoder and decoder in FP16 and keeps box decoding and reading-order postprocessing in FP32. The mask/polygon head is not included. See the source repository for the export recipe.

I/O

• Inputs: image float32 [1,3,800,800] RGB in [0,1] without mean/std normalization; im_shape float32 [1,2] = [800,800]; scale_factor float32 [1,2] = [1,1].
• Output: det float32 [300,7] = [classId, confidence, xmin, ymin, xmax, ymax, readingOrder]. Coordinates are in the 800×800 input space; scale_factor does not rescale them.
• 25 classes, same order as PP-DocLayoutV3.

Validation

Compared with the official FP32 PaddlePaddle/PP-DocLayoutV3_onnx (revision 46bbdf1) on 72 pages from four English papers, ONNX Runtime CPU, confidence ≥ 0.5 with PaddleX-style NMS:

Metric	Result
Regions	1434 vs 1434; 1431 matched at IoU ≥ 0.5
Mean IoU of matched regions	0.991
Label changes	0
Reading-order pairwise agreement	99.7%
Peak memory (process RSS)	0.77 GB vs 1.24 GB
CPU latency per page	unchanged (~1.1 s)

Remaining differences were regions with confidence at the 0.5 threshold. Scanned, CJK and vertical-text documents were not part of this comparison.

License

Apache-2.0, inherited from the upstream weights. See LICENSE.