Result: time 232s frames 7518; 32 FPS; 30 ms per frame

Accuracy: check the toyotamap.png

Note:

install

conda activate py37 (specific version required) conda install pytorch=1.3.1 cudatoolkit=10.0 -c pytorch (specific version required) pip install yacs pip install scikit-image pip install Pillow==6.1 (specific version required) conda install torchvision -c pytorch

realated https://github.com/prclibo/kitti_eval https://github.com/sshaoshuai/PointRCNN/tree/master/data/KITTI/ImageSets

Set up

#load the weight model from lacal path, change it in configs/smoke_gn_vector.yaml WEIGHT: "/home/carla/SMOKE/model_final.pth"

using carla to capture images and gps data

#For example: cd /opt/carla-simulator/PythonAPI/examples/examples python measurement_data_1.py python measurement_data_2.py the collcted data is stored in /home/carla/carla_data/

using python generate_location.py to generate grountruth and location difference file

mamunally set here, reason and challenges:

gps and image frames are in different frequency (1 frame with ~20 gps)

the start and end time of each vehicle is not sychronized (the ground truth are not 100% accurate because of the time shift)

camera parameters are different between the carla environments and existing trained models (camera location shift)

object detection results are not very accurate

prepare the images and test.txt file, follow the setting in generate_location.py

python modify.py

predict results

python tools/plain_train_net.py --eval-only --config-file "configs/smoke_gn_vector.yaml"

get evaluation result

python evaluate.py

matlab toyota_map.m for visualization

SikkeyHuang/SMOKE