Why I am getting different prediction, with tiny yolo.

[sagarkar10]

I am adding the images at two different predictions at two different conf_threshold but none of them is producing the desired results. The lower threshold image is added to show that the detection itself is distorted. conf_threshold=0.1

conf_threshold=0.5 (default)

@mystic123 Can you please tell me what I might be missing. weight file is yolov3-tiny.weights. I have also tried with letterbox reshape, but still, the detections are wrong. Anchors used are default anchors only. The image tested on is from darknet's official data folder.

[sagarkar10]

@mystic123 Can you please comment?

[mystic123]

Have you compared these results with detections from pjreddie’s implementation?

[sagarkar10]

Yes, I have, and the bicycle is quite accurate there...

[suryafyi]

@sagarkar10 , were you able to find a solution for this? I face the same problem with tiny-yolov3. Some detections are missing with the tf implementation.

[lucy3589]

@sagarkar10 , were you able to find a solution for this? I face the same problem with tiny-yolov3. Some detections are missing with the tf implementation.

hello, Could you slove it? I meet the same bug.

[sagarkar10]

For me ot was the bbox implementation and merging of bboxes and also the threaholds

[lucy3589]

For me ot was the bbox implementation and merging of bboxes and also the threaholds

Thanks for your reply. Could you give me your code about it. My address is [email protected]

[lucy3589]

For me ot was the bbox implementation and merging of bboxes and also the threaholds

what are your threadholds? I set some number, but can not see a pair of best threadholds.