AHP stands for Amodal Human Perception, which consists of 56,599 images from several large-scale instance segmentation and detection datasets (COCO, VOC (w/SBD), LIP, Objects365, OpenImages). Each image has a pixel-level complete human segmentation mask.

This dataset was originally proposed to solve the human de-occlusion task.

The dataset contains:

• Training set: contains a total of 56,302 images with comprehensive human annotations.
• Validation set: contains a total of 297 synthetic occluded images.
• Test set: contains 56 artificial occlusion images in total.