Skip to content

Latest commit

 

History

History
70 lines (50 loc) · 6.55 KB

MODEL_ZOO.md

File metadata and controls

70 lines (50 loc) · 6.55 KB
Raw

MODEL ZOO

Common settings and notes

  • The experiments are run with PyTorch 1.1, CUDA 10.0, and CUDNN 7.5.
  • The training is conducted on 4 V100 GPUs in a DGX server.
  • Testing times are measured on a TITAN Xp GPU with batch size 1.

nuScenes 3D Detection

We provide training / validation configurations, logs, pretrained models, and prediction files for all models in the paper

VoxelNet / SECOND / CBGS

Model Test time Validation MAP Validation NDS Link
cbgs_reimplemented 78ms 51.9 62.2 URL
centerpoint_voxel_1024 76ms 55.6 64.0 URL
centerpoint_voxel_1024_circle_nms 69ms 55.4 63.8 URL
centerpoint_voxel_1024_dcn_circle_nms 76ms 55.4 63.4 URL
centerpoint_voxel_1440_circle_nms 101ms 56.1 64.5 URL
centerpoint_voxel_1440_dcn_circle_nms 118ms 56.5 65.0 URL
centerpoint_voxel_1440_dcn_flip_circle_nms 440ms 58.8 66.9 URL
centerpoint_voxel_1440_dcn_flip_rotated_nms 449ms 59.1 67.1 URL

PointPillars

Model Test time Validation MAP Validation NDS Link
pointpillars_reimplemented 42ms 45.5 58.4 URL
centerpoint_pillar_512 41ms 48.3 59.1 URL
centerpoint_pillar_512_circle_nms 33ms 48.3 59.1 URL
centerpoint_pillar_512_dcn_circle_nms 41ms 48.6 59.4 URL

nuScenes 3D Tracking

Model Tracking time Total time Validation AMOTA ↑ Validation AMOTP ↓ Link
centerpoint_megvii_detection 1ms 192ms 59.8 0.682 URL
centerpoint_pillar_512_circle_nms 1ms 34ms 54.2 0.657 URL
centerpoint_voxel_1024_circle_nms 1ms 70ms 62.6 0.630 URL
centerpoint_voxel_1440_dcn_flip_circle_nms 1ms 441ms 65.5 0.586 URL
centerpoint_voxel_1440_dcn_flip_rotated_nms 1ms 451ms 65.9 0.567 URL
  • centerpoint_megvii_detection is our tracking algorithm with the public detection obtaiend from the nuScenes website

nuScenes test set Detection/Tracking

Our testset submission uses a single CenterPoint-Voxel model with flip test. It is an earlier version of our model that didn't use the Circle-NMS. Feel free to adapt our models to your framework and beat us on the leaderboard.

Detection

Model Test MAP Test NDS Link
centerpoint_voxel_1440_dcn_flip(Rotated NMS) 60.3 67.3 Detection

Tracking

Model Test AMOTA Test AMOTP Link
centerpoint_voxel_1440_dcn_flip(Rotated NMS) 63.8 0.555 Tracking

Notes

  • Training on 8 GPUs is OK, if the linear learning rate rule is applied.
  • We don't have result for 2 GPU / smaller batch size training. Though it should also work with linear learning rate rule.
  • If you face the out of cpu memory error for training, please reduce the num_worker
  • The io/CPU, and num_worker for the dataloader are crucial for the training speed. By default, we use 8 worker.
  • We observe up to 0.2 nuScenes map jittering due to the randomness of cudnn and fixed voxelization. It seems that turning on cudnn batchnorm gives slightly better testing speed and accuracy.