Skip to content

Latest commit

 

History

History
86 lines (66 loc) · 4.02 KB

MODEL_ZOO.md

File metadata and controls

86 lines (66 loc) · 4.02 KB
Raw

MODEL ZOO

Common settings and notes

  • The experiments are run with PyTorch 1.4.0, CUDA 10.1, and CUDNN 7.5.
  • The models can be downloaded directly from the download links.

Monocular 3D Detection / Tracking

nuScenes

Model GPUs Train time Val [email protected] Val AMOTA Download
nuScenes_3Dtracking 8 144h 0.352 0.242 model
nuScenes_LSTM_motion_model 1 9h - - model

Place the tracking model weight under

${QD-3DT}/work_dirs/Nusc/quasi_r101_dcn_3dmatch_multibranch_conv_dep_dim_cen_clsrot_sep_aug_confidence_scale_no_filter/latest.pth

Place the LSTM model weight under

${QD-3DT}/checkpoints/batch128_min10_seq10_dim7_VeloLSTM_nuscenes_100_linear.pth

Notes

  • Tracking model is trained on the keyframes of 6 camera images for 24 epochs on servers with 8x 32G V100 GPUs.
  • LSTM model is trained on the pure detection results and greedy matching to the groundtruth on servers with 1x 1080Ti GPU.

Waymo

! NOTE: Due to Waymo Open Dataset license, we cannot release the models trained on the dataset.
Model GPUs Train time Val MOTA/L2 [0m,30m)] Download
Waymo_3Dtracking 8 336h 0.0001 model
Waymo_LSTM_motion_model 1 24h - model

Place the tracking model weight under

${QD-3DT}/work_dirs/Waymo/quasi_r101_dcn_3dmatch_multibranch_conv_dep_dim_cen_clsrot_sep_aug_confidence_scale_no_filter_scaled_res/latest.pth

Place the LSTM model weight under

${QD-3DT}/checkpoints/batch128_min10_seq10_dim7_VeloLSTM_waymo_100_linear.pth

Notes

  • Tracking model is trained on all 5 camera images for 24 epochs on servers with 8x 32G V100 GPUs.
  • LSTM model is trained on the pure detection results and greedy matching to the groundtruth on servers with 1x 1080Ti GPU.

KITTI

Model GPUs Train time MOTA (2D Test) Download
KITTI_train_3Dtrack 4 16h 86.41 model
KITTI_LSTM_motion_model 1 0.5h - model
KITTI_subtrain_3Dtrack 4 14h - model
KITTI_subtrain_LSTM_motion_model 1 0.5h - model

Place the full train tracking model weight under

${QD-3DT}/work_dirs/KITTI/quasi_dla34_dcn_3dmatch_multibranch_conv_dep_dim_cen_clsrot_sep_aug_confidence_mod_anchor_ratio_small_strides_GTA/latest.pth

and half train tracking model weight under

${QD-3DT}/work_dirs/KITTI/quasi_dla34_dcn_3dmatch_multibranch_conv_dep_dim_cen_clsrot_sep_aug_confidence_subtrain_mod_anchor_ratio_small_strides_GTA/latest.pth

Place the LSTM model weight under

${QD-3DT}/checkpoints/batch8_min10_seq10_dim7_train_dla34_regress_pretrain_VeloLSTM_kitti_100_linear.pth

and

${QD-3DT}/checkpoints/batch8_min10_seq10_dim7_subtrain_dla34_regress_pretrain_VeloLSTM_kitti_100_linear.pth

Notes

  • Tracking model is trained on both detection and tracking images for 24 epochs on servers with 4x 32G V100 GPUs.
  • LSTM model is trained on the pure detection results and greedy matching to the groundtruth on servers with 1x 1080Ti GPU