Scaling YOLOv3 predictions instead of resize may have a better performance?

[xtyDoge]

The prediction of YOLOv3 usually a rectangle and I see that you use transforms method to resize the image before put it into HRNet. It stretching the image and make pose estimation network performs badly.
I try to make YOLOv3 prediction regions have the same h/w ratio as the input of HRNet. For example, HRNet requires image_resolution (256, 256)，but YOLOv3 gives a rectangle region. What we should do is calculate the center of YOLO prediction region and expand it as a square.

In SimpleHRNet._predict_single I modify it by adding:

squareLen = max(x2 - x1, y2 - y1) // 2
centerX = x1 + (x2 - x1) // 2
centerY = y1 + (y2 - y1) // 2
x1 = max(0, centerX - squareLen)
x2 = min(len(image[1]), centerX + squareLen)
y1 = max(0, centerY - squareLen)
y2 = min(len(image[0]), centerY + squareLen)

And it works on pose_hrnet_w32_256x256 with mpii annotation.
Maybe you can add it on your proj :D

[stefanopini]

That's a cool idea, thank you! I will surely add it to the project, probably as an option.

Have you tested also with models pre-trained on COCO and bounding boxes with aspect ratio 4/3 (as pose_hrnet_w32_256x192 and pose_hrnet_w48_384x288)?

[xtyDoge]

@stefanopini Sorry I don't, maybe you can try it.

[stefanopini]

Added from commit af40b39
Thank you for the suggestion!