search

hehefan / PSTNet2

Implementation of the "Deep Hierarchical Representation of Point Cloud Videos via Spatio-Temporal Decomposition" paper.
MIT License
15 stars 2 forks source link

Data Converters for N-UCLA and UWA3DII #2

Open jx-zhong-for-academic-purpose opened 2 years ago

jx-zhong-for-academic-purpose commented 2 years ago

Hi Hehe, your works about dynamic point clouds are so impressive, really inspiring me a lot 👍 👍 👍

I would like to conduct experiments on N-UCLA and UWA3DII as you do - for fair comparisons, if possible, could you provide the data converters for these two datasets?

Big thanks!

hehefan commented 2 years ago

Hi Jiaxing,

Thank you!

I am sorry I did not keep the datasets and the scripts. It is a bit complicated and messy to process the two datasets. Moreover, I do not think my processing is 100% correct.

Basically, you can use the https://github.com/hehefan/Point-Spatio-Temporal-Convolution/blob/main/scripts/depth2point4ntu.py script to convert the datasets. I did not find the intrinsic focal parameters and used the default 280 (potentially problematic).

In UWA3D, the gaps between performers and backgrounds are relatively obvious. I filtered most background points by setting depth thresholds (around 4000). Other non-person points are removed based on vertical and horizontal positions. You can check the filter results via mayavi visualization.

In N-UCLA, the dataset provides depth visualization images. I removed backgrounds according to them.

Best.

jx-zhong-for-academic-purpose commented 2 years ago

Many thanks for your prompt reply, Hehe!

Let me try to re-implement them :)