Inputs into PolarNet

[polars05]

PolarNet seems to take in a 9-dimensional feature vector as an input; could you clarify what these 9 features are?

Based on spherical_dataset in ./dataloader/dataset.py, my guesses are --

• return_xyz: (3 features) [**not too sure what this is**]
• xyz_pol: (3 features) original x, y and z coordinates converted to radius, angle and z, size [num_points, grid_size_r, grid_size_theta, grid_size_z]
• xy: (2 features) original x and y coordinates, size [num_points, 3]
• sig: (1 feature) reflectance, size [num_points, 1]

Are my guesses correct / wrong? Thanks!

[edwardzhou130]

return_xyz is the residual distances of points from the center of the assigned grid.

[polars05]

Thanks for your clarifications, @edwardzhou130! Why do you choose to input so many variations of spatial representation (return_xyz, xyz_pol, xy) instead of just, for example, xyz_pol and reflectance?

[YangZhang4065]

Hi @polars05 ,

Thanks for asking and very insightful questions. return_xyz (not even reflectance) is all you need if you only need to set up a decent baseline. We found additional seemingly-redundant xyz improve 0.5% mIoU (c.f. 9F in Table 5 in our paper). This trick is also used in SqueezeSeg papers and we are inspired by them.

Also, you can replace sig with your customized features (RGB + reflectance/ RBG only/ nothing at all).

[polars05]

Hi @YangZhang4065, it was a very interesting paper, and hence my curiosity / questions 😄

To clarify that my understanding of return_xyz ("residual distances of points from the center of the assigned grid" as mentioned by @edwardzhou130) is correct, is this what it means?:

Is so, why not convert the points and grid to polar coordinates first, and then use the residual distances in the polar space instead?

[edwardzhou130]

@polars05, sorry for the confusion caused by the name. I used the same name return_xyz in both voxel_dataset and spherical_dataset class, despite return_xyz in spherical_dataset is actually the distance in the polar space.

PolarSeg/dataloader/dataset.py

Lines 239 to 247 in 73488e8

	# center data on each voxel for PTnet
	voxel_centers = (grid_ind.astype(np.float32) + 0.5)*intervals + min_bound
	return_xyz = xyz_pol - voxel_centers
	return_xyz = np.concatenate((return_xyz,xyz_pol,xyz[:,:2]),axis = 1)
	if len(data) == 2:
	return_fea = return_xyz
	elif len(data) == 3:
	return_fea = np.concatenate((return_xyz,sig[...,np.newaxis]),axis = 1)

[polars05]

Thanks for clarifying my confusion! And residual distances refer to the l1, l2, l3, etc. in my above diagram?