Skip to content

Latest commit

 

History

History
149 lines (99 loc) · 3.62 KB

benchmark.md

File metadata and controls

149 lines (99 loc) · 3.62 KB
Raw

Benchmark

Mono 3D

Main results

methods dataset speed(fps) ap@0.7 ap@0.5 Download
geometry constrain kitti ~6 ~6 ~23 xxx
OFT kitti xxx xxx 10 xxx
keypoint 2d kitti ~6 xxx xxx xxx

Scripts

# keypoint 2d (use plane prediction)
# training
CUDA_VISIBLE_DEVICES=0 python train.py --cuda \
    --net fpn_corners_2d \
    --out_path /data/object/liangxiong/test \
    --config configs/fpn_corners_2d_mono_3d_kitti_config.json \
    --model /data/object/liangxiong/test/fpn_corners_3d/mono_3d_kitti/detector_600000.pth
# Note that to use --model option to load pretrained model
# (here to load pretrained 2d car detection model)
# And if you want to use multigpus just add more gpu ids
# to CUDA_VISIBLE_DEVICES and add --mGPUs option

# inference(checkpoint number)
CUDA_VISIBLE_DEVICES=1 python test.py --cuda \
    --checkpoint 600000 \
    --load_dir /data/object/liangxiong/test \
    --net fpn_corners_2d \
    --thresh 0.5 \
    --dataset nuscenes \
    --img_dir /data/dm202_3w/left_img \
    --calib_file ./000004.txt
# --calib_file refers to single calib file
# --calib_dir refers to directory of calibs(calib format is like that of kitti)
# --img_dir refers to directory where you want to infer

# inference(model path)
CUDA_VISIBLE_DEVICES=1 python test.py --cuda \
    --model ./faster_rcnn_32_3257.pth \
    --config ./configs/refine_kitti_config.json
    --net fpn_corners_2d \
    --thresh 0.5 \
    --dataset nuscenes
# note that if no dir of file is specified, use the val dataset to infer
# In the following snippet, the code of inference is omited.
# geometry constrain
CUDA_VISIBLE_DEVICES=0 python train.py --cuda \
    --net fpn_mono_3d \
    --out_path /data/object/liangxiong/test \
    --config configs/fpn_mono_3d_kitti_config.json \
    --model /data/object/liangxiong/fpn_bdd_pretrained/fpn/bdd/detector_300000.pth
# OFT

2D

Main results

methods dataset speed(fps) ap@0.7 Download
fpn_faster_rcnn kitti x ~86 xxx
faster_rcnn kitti xxx xxx xxx
faster_rcnn coco ~6 xxx xxx
faster_rcnn bdd ~6 xxx xxx
faster_rcnn nuscenes ~6 xxx xxx
ssd nuscenes xx xxx xxx
prnet nuscenes xx xxx xxx

Scripts

# KITTI
CUDA_VISIBLE_DEVICES=1 python train.py --cuda \
    --net fpn \
    --out_path /data/object/liangxiong/fpn_kitti_pretrained \
    --config configs/fpn_kitti_config.json
# BDD
CUDA_VISIBLE_DEVICES=1 python train.py --cuda \
    --net fpn \
    --out_path /data/object/liangxiong/fpn_bdd_pretrained \
    --config configs/fpn_bdd_config.json

CUDA_VISIBLE_DEVICES=1 python train.py --cuda \
    --net fpn \
    --out_path /data/object/liangxiong/fpn_coco_pretrained \
    --config configs/fpn_coco_config.json

Note that all model use the same backbone(res18_pruned)

PointCloud

ToolKits

Automatic generate config

python utils/generate_configs.py

If you just want to test the correction of your algorithm, enable DEBUG mode in file

Visualization

python utils/drawer.py

For each different dataset, just to uncomment the counterpart configs is Ok