habitat-lab
habitat-lab copied to clipboard
Objectnav datasets use `misc` category
🐛 Bug
The episodes generated for the mp3d objectnav
task include the misc (40)
item category. This likely hinders agent training, as misc
is too vague for even a human to understand. Current/future datasets would probably benefit from exclusion of misc (40)
semantic targets, as well as: objects (39)
, unlabeled (41)
, void (0)
, wall (1)
, floor(2)
.
Steps to Reproduce
Steps to reproduce the behavior:
cfg:
DATASET:
TYPE: ObjectNav-v1
SPLIT: val_mini
DATA_PATH: "/root/habitat-lab/data/datasets/objectnav/mp3d/v1/{split}/{split}.json.gz"
SCENES_DIR: "/root/habitat-lab/data/scene_datasets/"
code:
scene = self._env.sim.semantic_annotations()
label_to_str_cat = {
int(obj.id.split("_")[-1]): obj.category.name()
for obj in scene.objects
}
print(label_to_str_cat[obs['objectgoal'][0]])
>>> "misc"