i'm using yolov3.weights to test data/sample/, i found there are some wrong bboxes which are different from this projects

[J-LINC]

I am testing on/data/sample using the weights trained on the coco dataset, but the results obtained are somewhat different from those shown by the author on the homepage. That is why?

[Flova]

What weights did you use? Did you train from scratch or did you use the official darknet yolov3 ones?

[Flova]

Also make sure that the image size is 608 and not 416 (default).

[J-LINC]

What weights did you use? Did you train from scratch or did you use the official darknet yolov3 ones?

I'm not using my own training, it's the official training weight yolov3.weights

[J-LINC]

Also make sure that the image size is 608 and not 416 (default).

I really use 416 * 416, and I want to know if the model trained on 416 * 416 predicts that the image of 608 * 608 will work better, just like here.

[Flova]

I think the full size v3 is trained on 608 iirc..

[J-LINC]

I think the full size v3 is trained on 608 iirc..

I tried and the effect has improved, but there are still some problems. Do you think this is normal?

[Flova]

I think the errors are in levels that I expect from the v3 model. But I'm wondering why the boxes are slightly different now. It could be a numerical thing due to newer library versions etc., as I can not think of any significant changes to this part of the code. I also evaluated the weight on coco right now and got an mAP of 0.57653, which is slightly better compared to the value in the README.

[J-LINC]

I think the errors are in levels that I expect from the v3 model. But I'm wondering why the boxes are slightly different now. It could be a numerical thing due to newer library versions etc., as I can not think of any significant changes to this part of the code. I also evaluated the weight on coco right now and got an mAP of 0.57653, which is slightly better compared to the value in the README.

Okay,Thank you for your timely response！