The perception model currently being used is very simple. It extracts all of the vehicle objects from CARLA (essentially, the 'ground-truth') and assumes the vehicle is aware of all of them precisely.
There are two problems that fix_perception_json.py can help with:
- Sometimes, CARLA reports bogus bounding box extent values for bicycles (possibly for others too). If this program finds a bicycle, it replaces the bound box extent with a good one.
- Some scenarios are intended to demonstrate that the vehicle would not have been able to detect a vehicle in time to avoid a collision. For example, it might emerge suddenly from behind a fence. However, the ground-truth representation will suggest it was always visible. To help with this, a
--timestampargument can be provided. Any detections before this time (whether visible or not) will be removed.
cd ~/code/road/road-sim/apps
python3 fix_perception_json.py --input <logs_xxx.json-or-folder> \
--output <output-folder> [--timestamp <clip-timestamp>]
The output folder must be different to the input folder.
If --input specifies a folder, all JSON files in that folder will be processed. This is the typical use case.