stefanj0/PP-DocLayout-S-ONNX

PP-DocLayout-S — ONNX export

ONNX export of PaddlePaddle/PP-DocLayout-S, a lightweight (~4.7 MB) document-layout detector based on PicoDet/GFL, trained on a mixed Chinese/English corpus of papers, magazines, contracts, books, exams and research reports. 23 layout classes.

Hosted here for the RailReaderCore project — the detector intended for future web (WASM/ORT-Web) and mobile RailReader builds where the ~50 MB PP-DocLayoutV3 is too heavy.

Files

File	Size	Notes
pp_doclayout_s.onnx	4.7 MB	Exported via paddle2onnx from the upstream Paddle checkpoint.

Inference contract

The export keeps PaddleDetection's GFL-head two-input convention, not the three-input convention used by the PP-DocLayoutV3 RT-DETR export — so this is not a drop-in replacement at the I/O level.

Inputs

Name	dtype	Shape	Notes
image	float32	[1, 3, 480, 480]	Bilinear-resize the source image to exactly 480×480 without keep-ratio, scale to [0, 1], then ImageNet-normalise: mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225].
scale_factor	float32	[1, 2]	[480 / orig_h, 480 / orig_w] — the detection head divides predicted boxes by this, so the output boxes come back in the original image's pixel space.

Outputs

Name	dtype	Shape	Notes
Detections	float32	[M, 6]	Padded out to a fixed M. Each row is [class_id, score, x1, y1, x2, y2]. Coordinates are in the original image (un-resized) frame.
num_dets	int32/int64	scalar	Number of valid rows in the detection tensor. Trust this — the rest of the rows are padding.

NMS is baked into the graph at score_threshold=0.3.

Class table (23)

Ordered to match the upstream inference.yml:

 0  paragraph_title
 1  image
 2  text
 3  number
 4  abstract
 5  content
 6  figure_title
 7  formula
 8  table
 9  table_title
10  reference
11  doc_title
12  footnote
13  header
14  algorithm
15  footer
16  seal
17  chart_title
18  chart
19  formula_number
20  header_image
21  footer_image
22  aside_text

Minimal usage (ONNX Runtime, Python)

import numpy as np
import onnxruntime as ort
from PIL import Image

MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
STD  = np.array([0.229, 0.224, 0.225], dtype=np.float32)

sess = ort.InferenceSession("pp_doclayout_s.onnx", providers=["CPUExecutionProvider"])

img = Image.open("page.png").convert("RGB")
orig_w, orig_h = img.size
resized = img.resize((480, 480), Image.BILINEAR)
arr = (np.asarray(resized, dtype=np.float32) / 255.0 - MEAN) / STD
chw = np.transpose(arr, (2, 0, 1))[None]                         # [1, 3, 480, 480]
sf  = np.array([[480 / orig_h, 480 / orig_w]], dtype=np.float32) # [1, 2]

dets, num_dets = sess.run(None, {"image": chw, "scale_factor": sf})
for row in dets[: int(num_dets[0])]:
    cls_id, score, x1, y1, x2, y2 = row
    if score < 0.3:
        continue
    print(int(cls_id), float(score), float(x1), float(y1), float(x2), float(y2))

License

Apache-2.0, inherited from the upstream PaddlePaddle/PP-DocLayout-S.

Source verification

sha256:  33688dbee1c23e34b81777e97cb428eb40f24b242c02b5f623484959e830aec8
size:    4917852 bytes