The results of yolo v3 detector are different.

[DeHysolate]

I used the same cfg, weight and same image, but the results of running pjreddie / darknet in a Linux environment and AlexeyAB / darknet in a Windows environment are different.

I wonder what causes this result.

[XP-CHRC]

In ubuntu16.04, same cfg and weights, the AlexeyAB/darknet is better than the pjreddie/darknet, maybe you can check the function test_detector under the file detector.c, there are some differences.

[DeHysolate]

@XP-CHRC
I compared the original and AlexeyAB's, but found no significant change.
When the original and AlexeyAB's are using the same cfg and weight, does the result differ?

[AlexeyAB]

@DeHysolate

• Can you show whole command that you used?
• Check that you didn't specify -thresh in the command.
• Did you use default coco-model yolov3.weights or trained by yourself?
• Can you drag-n-drop this original image (without bound boxes) to next message here? I'll try to detect.

[DeHysolate]

@AlexeyAB

• Can you show whole command that you used?

detector test coco.data yolov3_train.cfg yolov3_train.weights save.jpg -i 0 -thresh 0.25

• Check that you didn't specify -thresh in the command.

thresh 0.25

• Did you use default coco-model yolov3.weights or trained by yourself?

myself based on yolov3.weight

• Can you drag-n-drop this original image (without bound boxes) to next message here? I'll try to detect.

Now I'm out and I can not capture it.

[AlexeyAB]

@DeHysolate Can you get different results on original / fork repositories using the default yolov3.weights model?

[DeHysolate]

@AlexeyAB I'll try it tomorrow and upload image because I'm out now. Thanks for your kindness!!

[DeHysolate]

@AlexeyAB
Result of using default weight (coco-model yolov3.weight)
The bounding box results are slightly different.
The probability results are also different. (origianl:100%,99% / fork:100%,97%)

[AlexeyAB]

@DeHysolate
Yes, there is a little difference due to different resize approaches: #232 (comment)