Experimental settings about multi_part_assembly

[Wuziyi616]

1. This is a good question and something to study. In our benchmark, I use the same training setting for all models on all categories (e.g. same learning rate, same epochs). And yes, I did observe overfitting on categories with a small number of data
2. Yes, the top one is training from scratch and the bottom one is doing finetuning, where you can see finetuning (transfer learning) leads to better results than training from scratch. We not only train 20 models for 20 categories, but we also have results where we train a single model over all 20 categories (see paper Table 11 in the last page). So we use this model trained on all categories for transfer learning.

[yaorz97]

Thank you for your prompt response!

[yaorz97]

What's more, I want to know if the model was trained in all categories in the 6.2 ablation study, or 20 models were trained with respect to each category.

[Wuziyi616]

I think that one is trained on each category separately. Actually I believe all the models in the main paper is trained on each category separately. We only put joint trained results in the appendix