in detection box[2] < box[0]

[ssusie]

Hi,

my training failed on epoch 13 when computing mAP the validation set. The detected boxes on an image have box[2] < box[0]. The detection looks like

[6.53528833e+00, 1.19281548e+02, 8.85683000e-01, 8.01131470e+02, 1.10095531e-01]

Any idea why this is happening and how I can fix it?

[hgaiser]

There are 5 values in that detection, what do they represent? Usually it is 4 values for the bbox, 1 value for the label and 1 value for the score.

[ssusie]

Thanks for the response!

I should have been more clear about that. The values I printed are from

all_detections, all_masks = _get_detections(generator, model, score_threshold=score_threshold,
                                                max_detections=max_detections, save_path=save_path)
detections = all_detections[i][label]

The first four coordinates represent a box, and the last one is the score of the box.

I also happen to have detection

[0.00000000e+00 7.02915344e+01 2.16546636e-02 8.10687439e+02
7.03362226e-02]

After converting this to int we get [0 70 0 810], which is not actually a box so the code fails on the line
mask = cv2.resize(mask, (box[2] - box[0], box[3] - box[1]))

[hgaiser]

Hmm strange, I've never seen this issue before. Are you training for just a few steps and then evaluating? Regardless, this shouldn't be an issue. If you want you can make a fix for the evaluation to ignore these boxes.