search

mit-han-lab / bevfusion

[ICRA'23] BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation
https://bevfusion.mit.edu
Apache License 2.0
2.38k stars 428 forks source link

Custom Dataset Implementation #568

Closed GoodarzMehr closed 7 months ago

GoodarzMehr commented 11 months ago

Hi, thanks for the great work!

For training BEV segmentation with a custom dataset I noticed that unlike with the nuScenes dataset where the forward direction (vehicle X axis) of model predictions and the ground truth were to the right, for me they had to point to the bottom (90 deg CW rotation), so I was wondering why. My data collection setup is pretty similar to nuScenes, one lidar, 6 cameras at the same angles and with the same resolution. The only difference is that unlike the nuScenes dataset where the lidar X direction points to the right, mine points forward (same as the vehicle X axis), so I wonder if that's been causing the issue.

baiguosummer commented 7 months ago

Hi, I read in the link you shared about the work on how to generate simulated nuscenes datasets using carla and simbev, I'm very interested in this, may I ask why the previous link to the repository is gone? @GoodarzMehr

GoodarzMehr commented 7 months ago

Hi, SimBEV originally only supported CARLA's Town06 map, but I recently made it private to work on expanding it to all CARLA maps. I have a working prototype and hopefully I'll make it public in a month or so.