Fast-BEV

Fast-BEV: A Fast and Strong Bird’s-Eye View Perception Baseline

Better Inference Implementation

Thanks to the repository CUDA-FastBEV inference using CUDA & TensorRT. And provide PTQ and QAT int8 quantization code. You can refer to it to get faster speed.

Usage

usage

Installation

• CUDA>=9.2
• GCC>=5.4
• Python>=3.6
• Pytorch>=1.8.1
• Torchvision>=0.9.1
• MMCV-full==1.4.0
• MMDetection==2.14.0
• MMSegmentation==0.14.1

Dataset preparation

  .
  ├── data
  │   └── nuscenes
  │       ├── maps
  │       ├── maps_bev_seg_gt_2class
  │       ├── nuscenes_infos_test_4d_interval3_max60.pkl
  │       ├── nuscenes_infos_train_4d_interval3_max60.pkl
  │       ├── nuscenes_infos_val_4d_interval3_max60.pkl
  │       ├── v1.0-test
  │       └── v1.0-trainval

download

Pretraining

  .
  ├── pretrained_models
  │   ├── cascade_mask_rcnn_r18_fpn_coco-mstrain_3x_20e_nuim_bbox_mAP_0.5110_segm_mAP_0.4070.pth
  │   ├── cascade_mask_rcnn_r34_fpn_coco-mstrain_3x_20e_nuim_bbox_mAP_0.5190_segm_mAP_0.4140.pth
  │   └── cascade_mask_rcnn_r50_fpn_coco-mstrain_3x_20e_nuim_bbox_mAP_0.5400_segm_mAP_0.4300.pth

download

Training

  .
  ├── work_dirs
    └── fastbev
      └── exp
          └── paper
              └── fastbev_m0_r18_s256x704_v200x200x4_c192_d2_f4
              │   ├── epoch_20.pth
              │   ├── latest.pth -> epoch_20.pth
              │   ├── log.eval.fastbev_m0_r18_s256x704_v200x200x4_c192_d2_f4.02062323.txt
              │   └── log.test.fastbev_m0_r18_s256x704_v200x200x4_c192_d2_f4.02062309.txt
              ├── fastbev_m1_r18_s320x880_v200x200x4_c192_d2_f4
              │   ├── epoch_20.pth
              │   ├── latest.pth -> epoch_20.pth
              │   ├── log.eval.fastbev_m1_r18_s320x880_v200x200x4_c192_d2_f4.02080000.txt
              │   └── log.test.fastbev_m1_r18_s320x880_v200x200x4_c192_d2_f4.02072346.txt
              ├── fastbev_m2_r34_s256x704_v200x200x4_c224_d4_f4
              │   ├── epoch_20.pth
              │   ├── latest.pth -> epoch_20.pth
              │   ├── log.eval.fastbev_m2_r34_s256x704_v200x200x4_c224_d4_f4.02080021.txt
              │   └── log.test.fastbev_m2_r34_s256x704_v200x200x4_c224_d4_f4.02080005.txt
              ├── fastbev_m4_r50_s320x880_v250x250x6_c256_d6_f4
              │   ├── epoch_20.pth
              │   ├── latest.pth -> epoch_20.pth
              │   ├── log.eval.fastbev_m4_r50_s320x880_v250x250x6_c256_d6_f4.02080021.txt
              │   └── log.test.fastbev_m4_r50_s320x880_v250x250x6_c256_d6_f4.02080005.txt
              └── fastbev_m5_r50_s512x1408_v250x250x6_c256_d6_f4
                  ├── epoch_20.pth
                  ├── latest.pth -> epoch_20.pth
                  ├── log.eval.fastbev_m5_r50_s512x1408_v250x250x6_c256_d6_f4.02080021.txt
                  └── log.test.fastbev_m5_r50_s512x1408_v250x250x6_c256_d6_f4.02080001.txt

download

Deployment

TODO

View Transformation Latency on device

2D-to-3D on CUDA & CPU

Citation

@article{li2023fast,
  title={Fast-BEV: A Fast and Strong Bird's-Eye View Perception Baseline},
  author={Li, Yangguang and Huang, Bin and Chen, Zeren and Cui, Yufeng and Liang, Feng and Shen, Mingzhu and Liu, Fenggang and Xie, Enze and Sheng, Lu and Ouyang, Wanli and others},
  journal={arXiv preprint arXiv:2301.12511},
  year={2023}
}