Open AgentEXPL opened 2 years ago
We consider two strategies for generating the ground-truth occupancy for anticipation. The one described in the paper in supplement section S5 will be used if GT_TYPE is set to grown_occupancy
. However, during our challenge submission (post the ECCV paper submission), we noticed that this strategy can cause the occupancy from other rooms to be included. These are typically very hard to predict since we may not have any visual context from an unseen room.
To deal with this, we used wall_occupancy
strategy for the challenge submission. It uses the the room walls to limit the growth of occupancy using a form of ray-casting. You can check out the implementation of these strategies here.
Great answer! But I was confused by wall_occupancy
for a long time and it cost me a lot of time ðŸ˜.
In the configuration file, the following is described for the GT_EGO_MAP_ANTICIPATED. What will happen if the GT_TYPE is selected as "wall_occupancy". Are there other types, and what are the meaning of the other types?
GT_EGO_MAP_ANTICIPATED: MAP_SIZE: 101 MAP_SCALE: 0.05 HEIGHT_THRESH: [0.2, 1.5] MAX_SENSOR_RANGE: 5.05 NUM_TOPDOWN_MAP_SAMPLE_POINTS: 20000 GT_TYPE: 'wall_occupancy' ALL_MAPS_INFO_PATH: 'data/datasets/exploration/gibson/v1/train/occant_gt_maps/all_maps_info.json'