Speedup FSDv2 via perception range shrinking

tusen-ai / SST

Code for a series of work in LiDAR perception, including SST (CVPR 22), FSD (NeurIPS 22), FSD++ (TPAMI 23), FSDv2, and CTRL (ICCV 23, oral).

Apache License 2.0

765 stars 96 forks source link

Speedup FSDv2 via perception range shrinking #160

Closed ArseniuML closed 6 months ago

ArseniuML commented 10 months ago

FSDv2 with pre-trained Argoverse2 weights seems to get good metrics on our dataset, but it would be great to speedup the detector by factor ~2-3. However, our Lidar has perception range +-100 m. Do I can speedup FSDv2 via perception range shrinking? Would it be sufficient to set perception range like this:

point_cloud_range = [-102.4, -102.4, -3.2, 102.4, 102.4, 3.2]
virtual_voxel_size=(0.4, 0.4, 0.4) #(512, 512, 16)

Abyssaledge commented 10 months ago

Yes, you can directly change the pc_range to this one. However, since FSDv2 is fully sparse, if there are no points beyond 100m in your dataset, I don't think it will be faster after shrinking the range. Anyway, you could have a try. If you have any problems after the modification, feel free to post new replies or issues.

ArseniuML commented 9 months ago

Perception range shrinking seems not to speedup the model. Are there other settings in config I can play with in order to speedup FSDv2?

Abyssaledge commented 9 months ago

The most straightforward way is to reduce the size of backbone, and the mixer. For example in this config, we adopt a very lightweight backbone, you could try it on FSDv2.