The thesis is developed by Yusuf Can Simsek. Note that this repository is designed to work in Artificial Intelligence and Robotics Laboratory at Politecnico di Milano (AIRLab) servers.
Docker environment
Expand
# Build the docker container which has the all components to run OSRM service, YOLO models and detectron2 models
docker build --rm -t <tag> -f Venv .
Clone the repository and download the model weights into the base folder. Then run bash script for the weights:
git clone https://github.com/yusiyohpolimi/thesis.git
cd traffic_lights_detection/
# Download the weights here and run following script
bash weights.sh
Map installation for OSRM service
Expand
cd osrm/
# To download all the germany map:
wget http://download.geofabrik.de/europe/germany-latest.osm.pbf
# For just Berlin:
wget http://download.geofabrik.de/europe/germany/berlin-latest.osm.pbf
# Extract the downloaded map:
run-docker --image_name="simsek:thesis_last" '' '' osrm-extract -p /osrm-backend/profiles/car.lua germany-latest.osm.pbf
# Partition:
run-docker --image_name="simsek:thesis_last" '' '' osrm-partition germany-latest.osrm
# Customize:
run-docker --image_name="simsek:thesis_last" '' '' osrm-customize germany-latest.osrm
# Do not forget to change the tag if you need
To train/test YOLO models, inputs should be in YOLO format. Check yolo_scripts for different cases. Note that the path for dataset is hardcoded. For multi-label task, use:
python3 yolo_format_bulb.py
CSP models (CSP10px, CSP640, CSP640multi):
Expand
train.py
--weights <path of the initial weights>
--cfg <model.yaml/.cfg path>
--data <data.yaml path>
--epochs
--batch-size
--img-size <single int input, e.g. 640>
--rect <rectangular training option for non square input sizes>
--resume <resuming from given weight>
--logdir <logging directory>
# check help for other arguments
test.py
--weights
--data
--batch-size
--img-size
--conf-thres <object confidence score threshold>
--iou-thres <IOU threshold for NMS>
--task <'val', 'test'>
--verbose <report results by class in bash>
# check help for other arguments
detect.py
--weights
--source <path of the input json file with image paths and vehicle data>
--output <folder to save detection outputs>
--img-size
--conf-thres
--iou-thres
# check help for other arguments
YOLOR models:
Expand
train.py
--weights <path of the initial weights>
--cfg <model.yaml/.cfg path>
--data <data.yaml path>
--epochs
--batch-size
--img-size <single int input, e.g. 1280>
--rect <rectangular training option for non square input sizes>
--resume <resuming from given weight>
--name <works like logdir in CSP models>
# check help for other arguments
test.py
--weights
--data
--batch-size
--img-size
--conf-thres <object confidence score threshold>
--iou-thres <IOU threshold for NMS>
--task <'val', 'test'>
--verbose <report results by class in bash>
--name <save folder for the results and plots>
# check help for other arguments
detect.py
--weights
--source <path of the input json file with image paths and vehicle data>
--name <folder to save detection outputs>
--img-size
--conf-thres
--iou-thres
# check help for other arguments
Detectron2 models:
Expand
train.py
--weights <path of the initial weights>
--data <json file of the trainset>
--epochs
--batch-size
--img-size <single int input, e.g. 1024 2048>
--resume <resuming from given weight>
--logdir <works like logdir in CSP models>
--anchors <using custom anchors generated>
# check help for other arguments
test.py
--weights
--data <json file of the validation/test set, use train json for validaiton task>
--img-size
--conf <object confidence score threshold>
--iou <IOU threshold for NMS>
--task <'val', 'test'>
--logdir <save folder for the results and plots>
--ratio <ratio of the test set to be used>
--anchors
# check help for other arguments
detect.py
--weights
--json-path <path of the input json file with image paths and vehicle data>
--output <folder to save detection outputs>
--img-size
--conf
--iou
--anchors
# check help for other arguments
CSP640multi:
# test
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> python3 csp/test.py --weights csp/weights/csp640multi/weights/best.pt --img 640 --data csp/data/dtld_bulb_100.yaml --iou 0.5 --conf 0.4 --batch 112 --task test --verbose
# detect
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> python3 csp/detect.py --weights ./csp/weights/csp640multi/weights/best.pt --img 640 --source berlin1_inp.json --iou 0.5 --conf 0.4 --output <output dir name>
YOLOR-D6multi:
# test
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> python3 yolor/test.py --weights yolor/weights/YOLOR-D6multi/weights/best.pt --img 640 --data yolor/data/ dtld_bulb_100.yaml --iou 0.5 --conf 0.4 --batch 12 --task test --verbose
# detect
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> python3 yolor/detect.py --weights yolor/weights/YOLOR-D6multi/weights/best.pt --img 2048 --source berlin1_inp.json --iou 0.5 --conf 0.4 --name <output dir name>
Detectron2:
# test
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> python3 detectron2/test.py --weights detectron2/weights/R50_30_130/model_best_ap50.pth --iou 0.0 --img 1024 2048 --conf 0.8 --data /multiverse/datasets/shared/DTLD/v2.0/v2.0/DTLD_test.json --anchors
# detect
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> python3 detectron2/detect.py --weights detectron/weights/R50_30_130/model_best_ap50.pth --img 1024 2048 --json berlin1_inp.json --iou 0.5 --conf 0.4 --output <output dir name> --anchors
Using the default_input.py script, given json file is cleaned such that there is no ground-truths. Then, new json file is generated to be used as input for the models.
python3 default_input.py
To run the full pipeline, use the bash script which first opens the OSRM service, and then starts pipeline.py which runs detection and mapping:
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> bash start.sh
# Check the start.sh for the given arguments of the pipeline.py.
It is also possible to run mapping on the detection json file:
run-docker --image_name="simsek:thesis_last" <gpu_number> <cpu_number> bash onlymapping.sh
# Check the onlymapping.sh for the given arguments of the mapping.py
- Building the docker image and extracting the map for OSRM takes a long-time. Both are needed for once.
- While running, snapping takes the most of the time after detection. Better GPS measurements will eliminate this.
- Disparity images only for Berlin are extracted since they consume quite a big space: /multiverse/datasets/shared/DTLD/Berlin_disp for Berlin1, /multiverse/datasets/shared/DTLD/Berlin_disp_full for all Berlin sequences.
- There are some hard coded paths for dataset location in multiverse, which is: /multiverse/datasets/shared/DTLD/
- Note that image_name should be the tag of the docker created in the first step.