search

Pointcept / PointTransformerV3

[CVPR'24 Oral] Official repository of Point Transformer V3 (PTv3)
MIT License
583 stars 30 forks source link

PTv3 Training Issues on large scale, density variable data #21

Open mihneawithh opened 3 months ago

mihneawithh commented 3 months ago

Hi, thank you for your great contribution. I am trying to train PTv3 for my data, which exhibits a large variety in density for a significant number of points (500,000+, indoor scenes). I trained both the PTv3 model and PointNet++ in parallel, and even though PTv3 was 2-3 times faster, it was not converging, while PointNet++ did. Could it be that PTv3 has difficulties with large variations in densities or high amount of points? Do you have any ideas what the problem could be and how we can use PTv3 for our data?

Gofinge commented 3 months ago

Hi, could you provide your training config?

mihneawithh commented 3 months ago

I used the base config parameters I did not modify any and grid_size of 0.07 (the same size as for the PointNet++ preprocessing)

Gofinge commented 3 months ago

Hi, could you provide detailed configs so I can go through and check whether some unexpected issues exist? It is quite strange that the model can not converge.