search

duanzhiihao / RAPiD

RAPiD: Rotation-Aware People Detection in Overhead Fisheye Images (CVPR 2020 Workshops)
http://vip.bu.edu/rapid/
Other
212 stars 63 forks source link

format of label when 'COCO' is used #15

Open kakusikun opened 4 years ago

kakusikun commented 4 years ago

Is the format of label same as the format used in COCO ? In paper, the model outputs the center of x and y, width, height, angle where the start axis is the y axis in cartesian coordinate and clockwise

if I have custom dataset, what is the format to used as COCO in this work?

duanzhiihao commented 4 years ago

When training on perspective (i.e., non-fisheye) images, the label format is the same as COCO. During training, angle will be added to each bounding box, and all the angle is assumed to be zero. https://github.com/duanzhiihao/RAPiD/blob/6c459010f86a614a5ecf7877506f4db7484fceff/datasets.py#L80