search

Marchetz / MANTRA-CVPR20

Official Pytorch code for MANTRA - Memory Augmented Neural Trajectory Predictor (CVPR2020)
Other
71 stars 15 forks source link

Regarding the generation of semantic map in BEV #12

Open JohnalDsouza opened 2 months ago

JohnalDsouza commented 2 months ago

Hi, could you let me know how the semantic maps were created.

fedebecat commented 1 month ago

Hi, it has been a while since we did that, but we used KITTI's transformations. I remember it not being too straightforward but you can project the segmentations onto the ground plane using KITTI's transformations (I believe there is a transformation for mapping the frames to the LIDAR's coordinate system, but I might be wrong). If I remember correctly, the main idea is that you can map segmentations to LIDAR and LIDAR to ground plane. If you do this for each frame you should be able to accumulate the projected segmentations into a whole top-view reconstruction of the scene of a a video. We then cropped the scene around the present instant of the agent we want to forecast, assuming it has limited visibility.

Again, it's been a while and this might not be super precise, but I hope it helps.