Thanks for your great implementation of sparse convolution and the paper~ I ran the 3D shape part segmentation experiments provided in examples/3d_segmentation recently and I modified the architecture according to what Benjamin pointed out in #10 . However, after training the network for 100 epochs, I'm getting around 83% mIoU in the 10-view test, instead of 86%, which is what you have achieved in the contest. Could you please help me find out the problem with the experiments? b.t.w. I modified the code so that the entire trainval set is used in training. Except for that, the data augmentation, evaluation scripts etc. are remain unchanged.
Hi authors,
Thanks for your great implementation of sparse convolution and the paper~ I ran the 3D shape part segmentation experiments provided in
examples/3d_segmentation
recently and I modified the architecture according to what Benjamin pointed out in #10 . However, after training the network for 100 epochs, I'm getting around 83% mIoU in the 10-view test, instead of 86%, which is what you have achieved in the contest. Could you please help me find out the problem with the experiments? b.t.w. I modified the code so that the entire trainval set is used in training. Except for that, the data augmentation, evaluation scripts etc. are remain unchanged.Best, Ken