AP=1.0 for all but one classes and no boxes drawn on test set

[giorgosrigas]

Hello guys. I am facing some issues on the testing phase: no boxes are drawn on the images and when I calculate the AP for each class for each image, three of my four classes always have AP = 1.0. Any idea as to why it could be happening? Thank you

[Sylv-Lej]

try lowering the bbox_threshold and debug using the initial list of probabilities of predicted bbxs

[charinho8]

Hi, i am facing the same kind of problem as mentioned above. Can you please explain a little further about the initial list of probabilities of predicted bbxs? Thanks in advance.

[giorgosrigas]

try lowering the bbox_threshold and debug using the initial list of probabilities of predicted bbxs

Thank you for the reply. Could you elaborate a bit more on the bit about the "list of probabilities of predicted bbxs"?

[Sylv-Lej]

probs is the list of predicted classes, you should see here some class and their probabilities.

if np.max(P_cls[0, ii, :]) < 0.7 or np.argmax(P_cls[0, ii, :]) == (P_cls.shape[2] - 1): continue

this line will exclude low probabilities or background predictions. By removing it you should be able to debug the predictions by showing probabilities of low confidence predictions. Debug using this then come back with your results

[Gerro385]

Hello, I also do not get any bboxes during testing and AP=1.0 on my 4 classes + bg

with the stated line active, I get no probs at all, with the line commented out, I get probs {'bg': [0.7346484, 0.7075835, ...]} does this mean my classes do net get recognized at all?

I also tried bbox_threshold = 0.01 and R = rpn_to_roi(Y1, Y2, C, K.image_dim_ordering(), overlap_thresh=0.01) but still no bboxes for me

what can i do? I did about 15 hours of cpu training with ~20k images

edit: I have overlapping objects edit: over night i trained another model with self.anchor_box_scales = [32, 64, 128, 256, 512] so i added two more options here and the training worked better in loss_rnp_regr and now i get some boxes in testing i can further work with