Who are like me: Fast human pose retrieval in unconstrained environments

Problems related to highly articulated humans are quite challenging in computer vision. The main difficulty lies in that a highly articulated human needs much more dimensions than a pedestrian-like human to cover all variations and situations. To cope with this difficulty, a dimension reduction strategy is required to convert the original problem into a tractable one. From such a point of view, we propose Adaptive Deformable Part based Model (ADPM) for the pose retrieval problem defined as retrieving similar human poses in large image datasets without annotations. ADPM involves two types of part models, static and dynamic. The static part models mainly describe parts with few variations, while the dynamic part models mainly describe parts with large variations, which supports our model to apply dimension reduction strategy. We predict human locations through a group of static parts and retrieve similar poses with a group of dynamic parts. Our ADPM, acting as a dimension reduction strategy, makes retrieving arbitrary poses possible. Experiments in unconstrained environments demonstrate the accuracy and efficiency of our approach.