Human Activity Knowledge Engine (HAKE) aims at promoting the human activity/action understanding. As a large-scale knowledge base, HAKE is built upon existing activity datasets, and supplies human instance action labels and corresponding body part level atomic action labels (Part States). We enhance the visual feature by combining it with the linguistic feature from BERT-Base pretrained model, which carries enormous contextual information from large-scale word corpus. With the power of HAKE, our baseline methods have outperformed state-of-the-art approaches on several wide-used activity and Human-Object Interaction benchmarks.
HAKE is still under construction, we will keep enriching and enlarging it. And we hope the volunteers from all the communities to work with us. We will make the dataset publicly available this summer.