To standardize the "public detection" evaluation protocol, we provide a pre-trained person detection model.
It is a Fully Convolutional Object Detection (FCOS) model that is trained on CrowdHuman train and validation dataset. It achieves 81.39% AP @ IOU = 0.5 AND 45.56% AP @ IOU = 0.75.
We encourage all researchers report their model's results with the provided detection results. The detection results can be downloaded with MOTChallenge format and GluonCV format.