Lane-Level Street Map Extraction from Aerial Imagery

This is the runable re-implementation code of paper Lane-Level Street Map Extraction from Aerial Imagery. This repo is UNofficial and is based on the offical LaneExtraction repo, but makes it runable, especially the inference scripts.

Notes

1. All experiments are conducted in the docker container. First go to ./docker to create the image and then run ./build_container.bash to run the container.

2. All experiments are conducted on a single GTX-1080Ti with python2. Since the source code of LaneExtraction is relatively old, do not use python3 or RTX3090 (RTX3090 may freaze the code due to CUDA issues). Before running the code, make sure you have selected the correct GPU in the bash scripts.

3. This repo only guarantees that it is runable and can produce reasonable results. But this repo is not well tested and the evalution part is not included. For any bugs, open issue in the official LaneExtraction repo or this repo.

4. For convenience, the pretrained checkpoints for inference are hardcoded written in the inference scripts. Refer to the inference scripts to change loaded checkpoints.

5. All inference outputs are stored in ./code/laneAndDirectionExtraction/output.

6. Find the discussion between me and the authors in issue #9.

Dataset

Run

./get_data.bash

Taining

cd laneAndDirectionExtraction
./run_train.bash

cd turningLaneValidation
./run_train.bash

cd turningLaneExtraction
./run_train.bash

Testing

1. Prepare pretrained checkpoints

./prepare_checkpoints.bash

cd laneAndDirectionExtraction
./run_test.bash

cd turningLaneExtraction
./run_test.bash