July 2020
tl;dr: A large scale (15k training images) dataset for crowded/dense human detection.
Very solid technical report from megvii (face++). Related datasets: WiderPerson.
Previous datasets are more likely to annotate crowd human as a whole ignored region, which cannot be counted as valid samples in training and evaluation.
- 22 human per image.
- Full body bbox (amodal), visible bbox (only visible region), head bbox. They are bound (associated) for each human instance.
- occlusion ratio can be quantified by the two bbox.
- evaluation metric:
- AP
- mMR (average log miss rate over FP per image)
- Image crawled from google image search engine, cleaned and annotated.
- Pervious datasets (CityPerson) annotates top of the head to the middle of the feet and generated a full bbox with fixed aspect ratio of 0.41.
- AP vs MR in object detection.