Evaluating pck

[amin-asdzdh]

When evaluating the PCK using the following function:

simple-HRNet/misc/utils.py

Lines 213 to 244 in f4a8174

	def evaluate_pck_accuracy(output, target, hm_type='gaussian', thr=0.5):
	"""
	Calculate accuracy according to PCK,
	but uses ground truth heatmap rather than y,x locations
	First value to be returned is average accuracy across 'idxs',
	followed by individual accuracies
	"""
	idx = list(range(output.shape[1]))
	if hm_type == 'gaussian':
	pred, _ = get_max_preds(output)
	target, _ = get_max_preds(target)
	h = output.shape[2]
	w = output.shape[3]
	norm = torch.ones((pred.shape[0], 2)) * torch.tensor([h, w],
	dtype=torch.float32) / 10 # Why they divide this by 10?
	norm = norm.to(output.device)
	else:
	raise NotImplementedError
	dists = calc_dists(pred, target, norm)
	acc = torch.zeros(len(idx)).to(dists.device)
	avg_acc = 0
	cnt = 0
	for i in range(len(idx)):
	acc[i] = dist_acc(dists[idx[i]], thr=thr)
	if acc[i] >= 0:
	avg_acc = avg_acc + acc[i]
	cnt += 1
	avg_acc = avg_acc / cnt if cnt != 0 else torch.tensor(0)
	return acc, avg_acc, cnt, pred, target

Why does the x_coordinate gets divided by the height and the y_coordinate by the width? Shouldn't it be the other way around?

Thanks

[stefanopini]

Hi! I've checked it and I think you're right.

This code is derived from here and the same code is found also here and they may both contain the same issue.

I've raised your question in microsoft/human-pose-estimation.pytorch#144 to have another check on this. I'll keep you posted.