Fix GUI detection colors by rendering the original image

[divyanshAgarwal123]

What changed

The image detection tab was converting the model input tensor back to a display image. When normalization is configured, that tensor no longer contains display-ready RGB values.

This change draws the predictions on the original uploaded RGB image instead. The prediction path already maps bounding boxes back to the original image coordinates. A regression test verifies that the original image is passed to the drawing function and that the normalized sample is not requested.

Reproduction

With a small RGB sample, the current conversion changed pixel (128, 64, 32) to (18, 23, 195).

Checks

• Black formatting check
• pytest: 47 passed

Fixes #524

[divyanshAgarwal123]

Hi @dpascualhe — I reproduced #524 on the current master branch and implemented the fix by drawing detections on the original uploaded RGB image, while keeping the existing mapping of predicted boxes to original-image coordinates. I also added a focused regression test, and all 47 local tests pass. The PR is ready for review when you have time. Thanks!

[divyanshAgarwal123]

Hi @dpascualhe, just following up on #631. It is still open and mergeable, and the focused regression test plus the full local suite (47 tests) pass. Please let me know if you’d like any changes or additional coverage. Thanks!