Here's a well-structured GitHub issue draft for your request: Issue: Generating Episodic Data for Object Navigation Challenge – Visibility Criteria

[sktalha68]

I am working on generating episodic data for the Object Navigation Challenge. Based on the paper, the required criteria for an episode include:

Agent’s initial location
Target object
Geodesic and Euclidean distances
The ratio of geodesic/Euclidean > 1.05
Valid viewpoints: Points within a 1m radius of the target object that are navigable and have visibility of the object

I am able to satisfy most of the criteria, but I am stuck on determining object visibility for a given viewpoint.
What I Have Tried

Semantic Sensor Data
    I used the semantic sensor from both HM3D Semantics and MP3D datasets.
    However, for some reason, the semantic sensor values for the HM3D  Sem dataset are all zeroes.
    To work around this, I tried using the semantic.glb file and extracting semantic information from the RGB sensor, which seems odd to me.

Mapping Pixel Values to Object Instances
    Even when I extract semantic information, the pixel values are too large (somewhere in an 1k-3k range).
    I couldn’t find a mapping between pixel values and object instances.

Questions

How do I correctly determine the visibility of an object from a given viewpoint?
Is my approach (using semantic sensor values) correct? If not, what is the recommended method?
Is there any reference code or structure used for episodic data generation?
If I am on the right track, how can I extract visibility information more reliably?

Expected Help

Guidance on correctly obtaining object visibility
Clarification on why semantic sensor values for HM3D appear as zeros
Any available codebase for episode generation to compare my approach

Any help would be greatly appreciated! Thanks in advance.