object detection dataset generator

This tool uses a deep learning tracker to generate bounding box annotations for detection datasets (COCO format).

Prerequisites

This tool requires pysot and the siammask_e add-on for pysot. Steps for installation are below.

1. Install requirements

   $ pip install -r requirements.txt

2. Clone pysot repo and follow instructions in the pysot INSTALL.md. The pip requirements above should cover everything needed for pysot so only the extension building step should be needed.
3. Clone SiamMask_E and run the install.sh script.
4. Download the siammask_r50_l3 .pth file from PySOT Model Zoo and place it in /path/to/pysot/experiments/siammaske_r50_l3/model.pth.
5. Export environment variables necessary for finding pysot:

   $ export PYTHONPATH=/path/to/pysot:$PYTHONPATH
   $ export PYSOTPATH=/path/to/pysot
   

Some usage examples

Generate bounding boxes for a video file. Set the dataset name and the class for the selected object. Images and a COCO json file will be saved in the results/ directory:

$ python coco_dataset_generator.py --video example_data/train/VID_20200730_151459.mp4 --dataset uticnice --class uticnica

Generate boxes for all videos in the selected directory and discard blurry frames:

$ python coco_dataset_generator.py --dir example_data/train/ --dataset uticnice --class uticnica --discard_blurry_frames

Manually select bounding boxes for all images in a directory:

$ python coco_dataset_generator.py --manual --dir example_data/valid/ --dataset uticnice_valid --class uticnica

If the selected bounding box was incorrect press 'c' to cancel and select it again. Tracking can be paused by pressing 'p'. Pressing 'r' during tracking resets the tracker - user selects a bbox for the current frame so the tracker can continue. Pressing 'n' ends the current video.

This project was tested with Python 3.7.6 and package versions found in requirements.txt