Fixed deduplication in picodet_postprocess

[trueMiskin]

Hi,

When I tested your OCR with layout analysis I ran into a weird bug. I had this image:

Firstly, I ran the model with a score threshold 0.4 and the output was this image:

I was a little bit upset that the model did not find the title with high confidence so I lowered the threshold to 0.3 and the result is this image:

When I saw it, I knew that something was going wrong. When the threshold is lowered, the same or more bboxes can appear, but they cannot disappear.

I thought there was some bug in the nms function but I found a bug in the deduplication function. The function iterates bboxes and computes iou with all other bboxes. If there are overlapping bboxes one persists and the other is removed (marked as duplicate) but this approach is wrong. For example in this case:

In the first loop iteration, bboxes 0, 1, and 3 are removed. This does not make any sense. Bboxes 1 and 3 do not overlap at all. Additionally these bboxes overlap with only bbox 0 which is removed in the same iteration as well. The correct approach is to remove only tested bbox 0.

After this fix, the output is as expected:

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

All committers have signed the CLA.

[Copilot]

Pull Request Overview

This PR fixes the deduplication process in picodet_postprocess by changing how duplicate bounding boxes are marked, ensuring that only the tested bbox is removed if it is not the highest scoring one.

• Simplifies the duplicate marking by checking if the current index differs from the kept index.
• Eliminates the unnecessary extension of duplicate indices using the overlaps list.

Comments suppressed due to low confidence (1)

ppocr/postprocess/picodet_postprocess.py:295

• The updated deduplication logic now only marks the currently tested bbox (i) as duplicate if it is not the highest scoring one (keep). Please double-check that this approach fully addresses the issue of incorrectly removing non-overlapping bboxes and that it does not inadvertently remove boxes that should be preserved.

if keep != i: