Processing text detection output

[Epigeos-com]

Hii, I'm running PP-OCRv6_medium_det_infer, through ONNX
This is my input:

This is the raw output:

Aand it looks pretty much correct, but moved and scaled, I'm not doing any preprocessing like that, and I couldn't find any info on how to get the original coordinates from this

[ngo-binh0108]

The output isn't moved and scaled by the model. The input tensor has the wrong memory layout: the image went in as HWC (interleaved RGB, the way OpenCV/PIL return it) without being transposed to the NCHW layout the model expects.

When HWC bytes are read as CHW, each of the three "channels" contains one third of the image rows, stretched 3× vertically and 3× horizontally. Your screenshots match that exactly. Your text is in the bottom third of the image, so it ends up in the third plane at y' = 3·y − 2·1088:

text line | y in input | predicted y' | y in your output -- | -- | -- | -- name | ~873 | 443 | ~437 subtitle | ~908 | 548 | ~550 main line | ~945 | 659 | ~660

Horizontally, the middle third of each row maps to x' = 3·x − 1920. That puts the name (x 910–1008) at 810–1104 (yours: ~830–1085) and the subtitle (880–1037) at 720–1191 (yours: ~748–1185).

The fix is to normalize and transpose before running the model. This matches PaddleOCR's own detection preprocessing:

import cv2
import numpy as np
img = cv2.imread(path)                       # BGR, HWC, uint8

x = img.astype(np.float32) / 255.0

x = (x - np.array([0.485, 0.456, 0.406], np.float32)) / np.array([0.229, 0.224, 0.225], np.float32)

x = x.transpose(2, 0, 1)[None]               # HWC -> NCHW, shape (1, 3, H, W)

prob = sess.run(None, {sess.get_inputs()[0].name: x})[0][0, 0]   # (H, W) probability map

With the correct layout, the probability map lines up with the input pixel for pixel. If you resize the image so both sides are multiples of 32, scale the boxes back by orig_w / input_w and orig_h / input_h.

[Epigeos-com]

It works! Thank youu