About vote option in test and noted data augmentation

[a18700]

About vote option in test and noted data augmentation

[yanx27]

@a18700 Hi, I actually use default setting in code for voting test. Voting exactly will introduce randomness to experiment results, but it just a slight change. Furthermore, voting test will not increase the performance and it works for more stable results, Finally, I add randomly drop points in trianing processing, which already used in official PointNet++ :)

[chenzhaiyu]

@ZJU-PLP @yanx27 I also encountered this issue (memory leaks) when setting num_workers > 0. The worker process will be killed after some iterations, especially because I don't have a large memory. This seems to be an issue introduced by the reference counting of Python. To mitigate it you can set num_workers = 0, which slows down your data IO for sure. See pytorch/pytorch#13246.

[a18700]

How kind you are!

One thing what i want to clarify remaining, how does the performance of pointnet++(ssg, msg) is better than official version?

For ssg, 92.4 and 90.9(in paper) is huge difference.

So at first i thought it's heavily depending on augmentations, but you replied as you continue to use same augmentations in paper.

Can you explain more about this enhancement in performance?

Thanks.

[yanx27]

@a18700 This work uses almost all the settings of the official version except:
(1) The ball query function maybe slightly different from official one
(2) I use both group normalized xyz and raw xyz to aggregate local group's features.
So I really do not sure that which one will enhance the performance :)

[ZJU-PLP]

@yanx27 Hi, could you mind sharing the memory usage when you train pointnet2_sem_seg? I find that the memeory need 20GB while offiical tensorflow version only 6GB!

Ps:the batch_size=16

[yanx27]

@ZJU-PLP Do you use same CUDA version and torch version with me?