About the Backbone

[kennethleng]

Hello author, I would like to ask during the eval process, after the range image is input into the model, what happens to the output y compared to the input range image? I shall: pr_img = delta_d * (delta_i).pow(3) > config.threshold pr_img &= xyz_img[:, 2, :, :] > config.z_ground pr_img &= range_img[:, 1, :, :] < i_thresh pr_img |= range_img[:, 0, :, :] < d_thresh change into pr_img = delta_d * (delta_i).pow(3) > 100000 pr_img &= xyz_img[:, 2, :, :] > 100000 pr_img &= range_img[:, 1, :, :] < 0.00001 pr_img |= range_img[:, 0, :, :] < 0.0001 It is found that the obtained results have no denoising effect, which means that the output y does not remove the noise. After the range image is input into the model, does the intensity value of the point cloud change? Thank you so much

[tn00364361]

what happens to the output y compared to the input range image

The output y ideally is the cleaned range image.

does the intensity value of the point cloud change

No.

Your threshold values are too extreme and do not match the physical properties of any real-world point clouds. Please read the comments in eval.py to understand how to pick reasonable d_thresh, i_thresh, and config.z_ground values.

As for config.threshold, it is a bit more complicated. You should generate scatter plots of delta_d v.s. delta_i for the entire dataset, and pick a curve that best separates snow points from non-snow points. For example, I picked delta_d * (delta_i).pow(3) = config.threshold as the dividing curve for this work.