This repository is the Python software toolkit for bounding-box visualization and algorithm evaluation on the CEPDOF dataset.
- [Oct 22, 2020]: Add instructions on evaluation on the HABBOF and MW-R dataset.
- [May 26, 2020]: Update docstrings and comments. Functionality is unchanged.
- [Apr 17, 2020]: Initial commit
CEPDOF's annotation format follows the COCO dataset convention, except that we use [cx,cy,w,h,degree (clockwise)] for each bounding box instead of [x1,y1,w,h] in COCO. /CEPDOF_sample is a toy sample of the CEPDOF dataset. For more details, please refer to our dataset page shown above .
Put the cepdof_api.py in your working directory and make use of the functions in it. Some examples are described below.
Requirements:
- Python3 (tested on Python 3.6 and 3.7)
- opencv-python
- pycocotools (for Windows users, please refer to this GitHub repo)
Example code for parsing and visualizing the annotations is provided in visualize_demo.ipynb.
Our evaluation code is built upon pycocotools so the usage is similar to it, except that we use [cx,cy,w,h,degree (clockwise)] instead of [x1,y1,w,h] for each bounding box. The detection results should be in the JSON format as in video_0_results.json. Example code for evaluation on CEPDOF is provided in eval_demo.ipynb.
Download HABBOF dataset, then convert the ground-truth labels to our JSON-format by running HABBOF_GtToJSON.py -p "<path_to_HABBOF>"
Follow the instructions here to download the dataset and annotations, then rename the frames with the renameMWImages.py script.
If you publish any work reporting results on the CEPDOF or the HABBOF dataset, please cite the corresponding paper.