how to create local data for training and what is the training label format?

[learn01one]

Hello, how to create local data for training and what is the training label format? , the hlw data cannot be downloaded and the data format cannot be known, thank you

[ebrach]

Hi. Sorry to hear that the HLW dataset is not accessible anymore. The request form (linked in our readme) still seems to exist - did you request access?

Anyway, another option to infer the dataset format is to reverse engineer hlw_dataset.py, specifically looking at

ngdsac_horizon/hlw_dataset.py

Lines 48 to 59 in 1bf0172

	for row in metadata:
	cur_image = 'hlw/images/' + row[0]
	if cur_image in self.images:
	idx = self.images.index(cur_image)
	cur_labels = torch.zeros((4))
	cur_labels[0] = float(row[1])
	cur_labels[1] = float(row[2])
	cur_labels[2] = float(row[3])
	cur_labels[3] = float(row[4])
	self.gt[idx] = cur_labels

and

ngdsac_horizon/hlw_dataset.py

Lines 73 to 84 in 1bf0172

	#convert ground truth coordinate system (from zero center to zero corner)
	yOffset = image.shape[0] / label_scale / 2 # mind zero padding to make image square
	xOffset = image.shape[1] / label_scale / 2
	gt = self.gt[idx].clone()
	gt /= label_scale
	gt[0] += xOffset
	gt[1] *= -1
	gt[1] += yOffset
	gt[2] += xOffset
	gt[3] *= -1
	gt[3] += yOffset

.

It looks to me like the metadata.csv of the dataset contained one line per image, as img_file, x0, y0, x1, y1. The line coordinates (start and end point) seem to be in px, and relative to the image center. y-direction seems to be from bottom to top.

But all of the above is a guess based on a quick look and has to be tested.

Best,
Eric