CrowdHuman dataset prepared for darknet implementation
Optional annotations were ignored, only visible boxes are used.
All annotations are standartized to <object-class> <x> <y> <width> <height>, where:
<object-class>- string class of object from (person or mask)<x> <y> <width> <height>- float values relative to width and height of image, it can be set from 0.0 to 1.0- for example:
<x> = <absolute_x> / <image_width>or<height> = <absolute_height> / <image_height> - attention:
<x> <y>- is center of rectangle (not top left corner)
Annotation files are included in /annotations folder.
Images are not part of this repo, please download it by yourself and put to matching folders:
/images_{suffix}, where {suffix} is test, train or val
Algorithm creates .txt annotation file for each image (matching ID in each annotation row)
Use python3 main.py to run script
Use pip3 install -r requirements.txt to install python dependencies