Mismatch Between Gaze360 Annotations and Downloaded Dataset Structure

[huangyctw]

We're re-implementing the method using this repo and encountered a mismatch between the provided Gaze360 annotations and the dataset downloaded from the official Gaze360 site. The original dataset stores repeated head/body images in a shared folder structure. Your annotations reference different filenames, including paths with Face/, which do not exist in the original structure.

Questions:

1. How are the annotation filenames mapped to the official Gaze360 dataset?
2. Is there a preprocessing step or script that restructures the dataset?
3. Did we miss any dataset preparation details not mentioned in the paper?

The paper (p.4439) only explains how 3D gaze vectors are converted to 2D, but doesn’t clarify this filename/folder mapping.

Thanks in advance.

[thohemp]

We used the preprocessing from this site: https://phi-ai.buaa.edu.cn/Gazehub/3D-dataset/
Following their procedure should lead to the same output structure as used in this work, shown here: https://phi-ai.buaa.edu.cn/Gazehub/Guideline/FaceBased/Gaze360.pdf
I hope this helps.