Closed sai-prasanna closed 1 year ago
Hi Sai, Thank you very much for your interest in the MultiON Challenge 2023! In our challenge, the semantic sensor is not available during test. Anyway, we prepared it for training purposes with the lines of code you pointed out. Given that, in the challenge, the semantic annotations for the meshes are not needed, we used the IDs from 0 to n, where n is the length of self.object_to_datset_mapping. If, during your training, you need a different representation, you can change the code locally as you prefer. Instead, during test, the semantic sensor is not available. Thus, having different representations for xxx.semantic_id won't affect your results.
Hi,
In this line you set the
semantic_id
property to the object's semantic id. But habitat uses instance ids in it'ssemantic_id
property, so directly setting semantic id here will set wrong index. If someone ends up using semantic mapping to solve the task, it won't work.https://github.com/3dlg-hcvc/multion-challenge/blob/3b016381ca8d833ac4d23dc7d20ee3111fa3930b/habitat/tasks/nav/multi_object_nav_task.py#L243
I have fixed this in my research project which uses your dataset, but unfortunately I change the task definition, simulator version (newest habitat). So I can't directly put a pull request. I will attach my changes, you can look it through.