Questions on the split files

[chrockey]

Hi,

Some datasets like 3RScan, ARKitScenes, HM3D, and MultiScan consist of multiple split files, each denoting different subsets. Specifically, they have three types of split files: {train/val}_split_before_filter.txt, {train/val}_split_non_overlap.txt, and {train/val}_split.txt. Could you clarify the distinctions among these split files?

Furthermore, in ProcThor, there are {train/val}_split_all.txt as well as {train/val}_split.txt. What is the difference between these files?

Thanks!

[yixchen]

Hi,

Sorry for the confusion. The {train/val}_split_non_overlap.txt is meant for the per-dataset experiment in the early stage of the project. Later we merge them to create a larger training set, i.e., train_split_before_filter.txt. In our final experiments, we use {train/val}_split.txt for all the datasets by filtering out 'problematic' scenes, e.g., too large, too small, or multi-floor. In short words,

1. Use {train/val}_split.txt to reproduce the experiments in the paper. The val.txt for MultiScan is the SceneVerse-val. There is no val.txt for ARKitScenes, HM3D, 3RScan because we are not validating on them.
2. The train_split_before_filter.txt = train_split_non_overlap.txt + train_split_non_overlap.txt.
3. train.txt + filter_out scenes = train_split_before_filter.txt
4. In ProcThor, {train/val}.txt + filter_out scenes = {train/val}_split_all.txt

We may clear out the unnecessary files if people find them confusing in a later version.

Best,

[chrockey]

Thank you for the reply! It resolved most of my confusion except train_split_all.txt. What is the difference between train_split_all.txt and train_split.txt in ProcThor? Is train_split_all.txt similar to train_split_before_filter.txt?

[yixchen]

Yes, you are right. And I have updated the previous comment accordingly.

[chrockey]

I see. Thank you very much for kind reply :)