[PathDet] Data curation - BDD100k

[sarun-hub]

Description

Submitting this PR for issue #8.

Format

Format is same as Tran's TuSimple.

Feature

1. Usage

python process_bdd100k.py. --mask_dir /path/to/drivable/path/mask --image_dir path/to/raw/image --output_dir path/to/output

2. Workflow

Cropping with crop margins.

• test

{
    "top" : 0,
    "right" : 140,
    "bottom" : 220,
    "left" : 140
}

3. Audit Data

Use lane detection marking (in Json). Filter the drivable path image, and drivable path json to new directory.
Note: need to do manually auditing again.

Conditions:

1. laneDirection is not vertical.
2. There is lane detected.
3. There is drivable path detected.
4. category is not crosswalk.

Can be done by executing audit_data.py with command line arguments.

• --lane_json: Path to lane detection json.
• --mask_dir: Path to drivable path mask directory.
• --output_dir : Path to post-auditing directory.
• --max_files: Maximum number of files to process.