search

liujia99 / CrossNet

[TMM 2023] Self-Supervised Intra-Modal and Cross-Modal Contrastive Learning for Point Cloud Understanding
7 stars 1 forks source link

The gap between the reproduction and your data is very large. #1

Open YangParky opened 6 months ago

YangParky commented 6 months ago

The experiments we run ourselves cannot replicate your experimental data. Can you please provide your experimental logs?

image

liujia99 commented 6 months ago

Thank you for your interest in this work CrossNet. Unfortunately, due to the passage of time, I did not keep the stored files from that time.

Based on your feedback, I guess there is some coincidence in this situation, because the images introduced during the training process are random. It is recommended that you set the batch size to 200, and pay attention to the setting of the temperature parameter in comparative learning. Different values should be set for different data sets.

The effect of CrossNet seems to be eliminated in current self-supervised learning methods. We recommend our other point cloud self-supervised work Inter-MAE, and CrossNet is explained and applied in more detail.