About point cloud generation task's parameters

[EchoTHChen]

For point cloud generation task of this paper, I want to confirm the parameters.

Are the networks trained under those parameters which are different from the code?

[EchoTHChen]

And I wonder what the chamfer distance results of the networks are?

[siyuan2018]

We used a FoldingNet-like network, which is slightly different from the original FoldingNet paper, and similar to the architecture in paper Real-time Soft Body 3D Proprioception via Deep Vision-based Sensing.

For the chamfer distance, the training loss is around 0.04 in the last iterations after converging.

[EchoTHChen]

Thanks! I find the point clouds are just be divided by max_range. If i want to use mean value or mid value to normalize point clouds, how point clouds should be normalized to [0,1] or [-0.5, 0.5] for all axes in data_collect.py file?

[siyuan2018]

Hi, I think one way you can try is, after you normalize with mean or mid value, check the range of all coordinates, then do the normalization again.