RPMNet
RPMNet copied to clipboard
yewzijian
Shuffle Problem
Hi, when I test the model with my own dataset, I found that the random seed for the shuffle transformation will strongly influence the performance of the model. Can you tell me more about how to set the ranom seed? I saw in your dataset, you just use the index of the data in the batch to be the random seed, when using the random sample transformation.
Hm, this is weird. While I'm unable to test it right now, it shouldn't significantly affect the results. The reason I set it to the index of the data is to ensure the test set is deterministic and reproducible, while still having a different sampling/transform/crop for each test point cloud.
Ok, thanks a lot. I will do more test to ensure that. Meanwhile, I have another problem is that the setting of the radius, one of the hyperparamters in the model. The 0.3 seems not suitable for all the mesh outside the modelnet40 dataset. Therefore, is there any suggestion about choosing a correct radius for the specific meshs?
Setting radius is a bit tricky. Too small will mean it doesn't contain enough information for the descriptor to be descriptive. Too large will affect its localization accuracy and underperform under partial occlusion. Also, the PointNet based network only samples a maximum of 64 points within the radius.
You need to take these in mind when you set. I recommend just trying different numbers and choose the one that works best.