What does "point cloud distillation loss" mean for CroCo V2 initialization?

[lizhiqi49]

Nice work! I have a question about model training. In appendix it is mentioned that "for CroCoV2 initialization, we warm up the training by adding point cloud distillation loss from the DUSt3R model for 1,000 steps". However, I checked the DUSt3R paper but did not see a "point cloud distillation loss" term. So I wonder what does this loss exactly mean?

[botaoye]

Hi, thanks for your interest in our work. The distillation loss is used by us instead of DUSt3R. You can find it here, it is similar with the point loss used by DUSt3R.

[sfchng]

Is there any change of other hyperparameters to run this point cloud distillation loss apart from setting self.train_cfg.distill_max_steps=1000?