Object-Goal-Navigation
Object-Goal-Navigation copied to clipboard
Running on Matterport3D
Hi,
Thanks for your amazing works. I am trying to get it run on Matterport3D, but an info file (train_info.pbz2) is required. Would you be able to provide the codes for generating that file?
Thanks!
I am looking for the info file for MP3D, too.
Thanks!
In my understanding that file has information useful to estimate the object-nav metrics. If you can use directly the simulator instead, then you can ignore the train_info.pbz2
In my understanding that file has information useful to estimate the object-nav metrics. If you can use directly the simulator instead, then you can ignore the train_info.pbz2
Yes, I directly used the simulator. But the problem here is how to determine the floor height, as one floor can span various heights in the MP3D dataset, and the floor height given by the level AABB annotation is unreliable.
In my understanding that file has information useful to estimate the object-nav metrics. If you can use directly the simulator instead, then you can ignore the train_info.pbz2
Yes, I directly used the simulator. But the problem here is how to determine the floor height, as one floor can span various heights in the MP3D dataset, and the floor height given by the level AABB annotation is unreliable.
I am also facing the same problem now. Is there an effective solution? Can we generate this info by ourselves
I don't think you need to generate this info if you are running experiments on MP3D. In get_metrics() at envs/habitat/objectgoal_env.py you can use directly the simulator to get the agent's position and distance from the goal positions. The goal positions can be found from the v1 dataset provided in the habitat-lab repository.
You can verify this by tracing where the train_info.pbz2 info are being used in objectgoal_env.py. You can remove any use of that in the code. Also you don't need to pre-determine the floor-heights. If your input episodes contain start_position, start_rotation, object_category, and starting_distance (geodesic) then you should be fine.
I don't think you need to generate this info if you are running experiments on MP3D. In get_metrics() at envs/habitat/objectgoal_env.py you can use directly the simulator to get the agent's position and distance from the goal positions. The goal positions can be found from the v1 dataset provided in the habitat-lab repository.
You can verify this by tracing where the train_info.pbz2 info are being used in objectgoal_env.py. You can remove any use of that in the code. Also you don't need to pre-determine the floor-heights. If your input episodes contain start_position, start_rotation, object_category, and starting_distance (geodesic) then you should be fine.
I see. I can use the v1 dataset directly. And this is a convenient way. Thank you!
Hi. I am also facing the similar problem. I wonder how to reset the episode without train_info.pbz2 when using mp3d dataset? For gibson, we use ground truth sem_map
from train_info.pbz2 to pick the new object in the new episode (function generate_new_episode), but how could this be done in mp3d?
Any suggestion would help! Thank you :)
train_info Hi Min Zhao, would you mind giving more details about your solution? How did you make it work on Mp3D? It would be very helpful to me. Thanks so much!
Could you give more details about your solution? How did you make it work on Mp3D? It would be very helpful to me. Thanks so much!