Train / test set split clarification

[hongsukchoi]

Thank you for the awesome dataset!

The description about the train/test set splits is pretty unclear in the paper and I couldn't find more information in this repository too.

We split the dataset into 77260 egocentric images for train and 47740 for test. The split ensures non-overlapping locations between sets.

According to the benchmark annotation file I downloaded, it seems you are using all 01_tagging, 02_lego, and 03_fencing sequences are the test dat and the rest of the data is the training data. So, the training data and testing data are only captured in different physical locations but also contain different types of motion. Am I right?

Best

[hongsukchoi]

• What are the manual superviison and manual error fixes described here? Thanks a lot!!!!

We further refine these 2D bbox proposals using FasterRCNN [97] and manual supervision. Lastly, we annotate the 2D human poses in a top-down fashion for all the views using HRNet-WholeBody [45, 112] along with manual error fixes (c.f., Fig. 2c).