Independent PyTorch-YOLOv3 Results

[glenn-jocher]

I tested the 3 box regression methods below on https://github.com/ultralytics/yolov3 using yolov3-spp.cfg with swish trained on full COCO2014 to 27 epochs each, but was not able to realize performance improvements with the new methods. I'll try again with LeakyReLU(0.1). The IoU function I implemented is here.

python3 train.py --weights '' --epochs 27 --batch-size 16 --accumulate 4 --prebias --cfg cfg/yolov3s.cfg

	mAP@0.5	mAP0.5:0.95	Epoch time on 2080Ti
GIoU (default)	49.7	30.2	36min
DIoU	49.4	30.0	36min
CIoU	49.7	30.1	36min

[glenn-jocher]

Here are the training plots for the 3. results73, results74 and results75 are GIoU, DIoU and CIoU.

[Zzh-tju]

1. I'm not sure about your trade-off weight of regression loss balanced with the other losses.
2. The definition of CIoU is a little different, note alpha.
3. Bbox regression may degenerate if applied for multiple iters. For example, I usually choose 49k and 50k weight files to test on VOC dataset. Wish these help you.

[glenn-jocher]

@Zzh-tju thanks for the feedback!

1. Yes we use loss balancing among the 3 losses (box, objectness, classification). We derived these from hyperparameter studies. The balancing we found lets us train to higher mAP than darknet on COCO (+4.5 mAP@0.5:0.95 at 416 resolution) see https://github.com/ultralytics/yolov3/issues/310#issuecomment-549629973.

2. Yes, I tried to optimize the implementation a bit by inlining the alpha parameter into the equation:

   v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
   return iou - (rho2 / c2 + v ** 2 / (1 - iou + v))  # CIoU

   rather than:

   v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / h2) - torch.atan(w1 / h1), 2)
   alpha = v / (1 - iou + v)
   return iou - (rho2 / c2 + v * alpha)  # CIoU

3. I don't understand. Can you explain please? For our training comparison we used full COCO2014 trainval (117,000 images) for 27 epochs (10% of full training time).

[Zzh-tju]

2. Alpha actually does not have a gradient backward, just as an adaptively changing coefficient. We have noticed that this is not clearly expressed in the paper.

3. The degeneration of regression can be referred to IoU-Net (https://arxiv.org/abs/1807.11590). And this phenomenon is very common in the detection pipeline using bbox regression. By the way, the timing of the maximum performance point is also uncertain.

[glenn-jocher]

@Zzh-tju ah so the alpha should be under a with torch.no_grad(): statement? Did you get different results this way?

[Zzh-tju]

yes