Inferene on my own dataset

[starrah]

Inferene on my own dataset

[Starrah]

Hello!
The .txt files contain information about the ground truth corners. In evaluation mode, we need the .txt files to calculate metrics like 3DIoU.
Of course, you may not need the files if you just want to inference on your images, without calculating metrics.
In such situation, you have two methods:

First, you may modify the code.
Remove the lines that read .txt files, so that self.cors keeps equals to []

DMH-Net/perspective_dataset.py

Lines 68 to 71 in 2d2a346

	for filename in self.img_fnames:
	with open(os.path.join(self.cor_dir, filename[:-4] + ".txt")) as f:
	cor = np.array([line.strip().split() for line in f if line.strip()], np.float32)
	self.cors.append(cor)

Change the line to return self.getItem(self.img_fnames[idx], None)

DMH-Net/perspective_dataset.py

Line 101 in 2d2a346

return self.getItem(self.img_fnames[idx], self.cors[idx])

And the most important, modify the getItem function,

DMH-Net/perspective_dataset.py

Line 103 in 2d2a346

def getItem(self, filename, cor=None):

remove all code lines that requires the cor variable to calculate. Those may include:

DMH-Net/perspective_dataset.py

Lines 325 to 370 in 2d2a346

	if self.cfg.MODEL.HOUGH.CLINE_TYPE == "NEW":
	liness.append(np.array(d["lines"]))
	try:
	peaks = self.linesToPeaksNew(d["lines"], self.view_size)
	except:
	assert False, AUG_RECORD
	peakss.append(peaks)
	peaks_for_label = peaks
	label_hw = self.view_size
	xPeaks, yPeaks, cUpPeaks, cDownPeaks = peaks_for_label
	xLabels.append(
	self.generate_gradual_hough_label(xPeaks, label_hw[1], type=self.hough_label_gradual_type,
	base=self.cfg.MODEL.HOUGH.GRADUAL_LABEL.XY))
	yLabels.append(
	self.generate_gradual_hough_label(yPeaks, label_hw[0], type=self.hough_label_gradual_type,
	base=self.cfg.MODEL.HOUGH.GRADUAL_LABEL.XY))
	cline_angle_num = label_hw[1] + label_hw[0] // 2 * 2 - 2
	cUpLabels.append(
	self.generate_gradual_hough_label(cUpPeaks, cline_angle_num, type=self.hough_label_gradual_type,
	base=self.cfg.MODEL.HOUGH.GRADUAL_LABEL.CUPDOWN))
	cDownLabels.append(
	self.generate_gradual_hough_label(cDownPeaks, cline_angle_num, type=self.hough_label_gradual_type,
	base=self.cfg.MODEL.HOUGH.GRADUAL_LABEL.CUPDOWN))
	else:
	raise NotImplementedError()
	n_cor = len(cor)
	gt_floor_coor = cor[1::2]
	gt_ceil_coor = cor[0::2]
	gt_floor_xyz = np.hstack([
	post_proc.np_coor2xy(gt_floor_coor, self.ch, self.W, self.H, floorW=1, floorH=1),
	np.zeros((n_cor // 2, 1)) + self.ch,
	])
	gt_c = np.sqrt((gt_floor_xyz[:, :2] ** 2).sum(1))
	gt_v2 = post_proc.np_coory2v(gt_ceil_coor[:, 1], self.H)
	gt_ceil_z = gt_c * np.tan(gt_v2)
	height = np.array([gt_ceil_z.mean() - self.ch], dtype=np.float32)
	# Convert all data to tensor
	e_img = torch.FloatTensor(e_img.transpose([2, 0, 1]))
	# angle = torch.FloatTensor(angle)
	# up_bin256 = torch.FloatTensor(up_bin256.copy())
	# down_bin256 = torch.FloatTensor(down_bin256.copy())
	height = torch.FloatTensor(height)

and delete the affected keys from the returned dict, which may include cor, height, *Labels, lines, peaks

DMH-Net/perspective_dataset.py

Lines 372 to 385 in 2d2a346

	out_dict = {
	"filename": filename,
	"e_img": e_img,
	"cor": cor,
	"height": height,
	"p_imgs": torch.stack(p_imgs, 0),
	"xLabels": np.array(xLabels).astype(np.float32),
	"yLabels": np.array(yLabels).astype(np.float32),
	"cUpLabels": np.array(cUpLabels).astype(np.float32),
	"cDownLabels": np.array(cDownLabels).astype(np.float32),
	"peaks": peakss,
	"lines": liness
	}
	return out_dict

And, change this line to r = {}

DMH-Net/e2plabel/e2plabelconvert.py

Line 44 in 2d2a346

r = generateOnePerspectiveLabel(e_img, cor, *view, VIEW_SIZE)

I think those may enough, but I cannot guarantee it. The principal is to remove all place that use the cor variable, so if you meets with any error about not finding the cor, just feel free to remove it.

Or, you may have another option to make pesudo .txt files. The example is:

85 175
85 349
376 142
376 383
710 186
710 336
878 196
878 325

In which each lines represents the coordinate of a corner point in the image. Note that if you make pesudo .txt files, the metrics output of our code is totally meaningless.

[ghvision]

Thank you for your kind reply.

I modified the code and it works fine!