Visual pattern discovery refers to capture re-occurring spatial structures of visual primitives, which have extensive applications in image analysis, such as object recognition, pedestrian search. For human-centric recognition task, the identification always composed of appearance attributes which lack of visual geometric correlations among body parts (e.g. right ankle in front of torso, left wrist is closer to head than right wrist). We explore to mine the semantic pose invariance called pose-aware pattern and get a universal understanding schema. And we propose a novel framework to discover pose-aware patterns and make use the higher-level cognition for human pose estimation. Concretely, the framework employs the notion of semantic pose consistency to quantify and generate external knowledge as visual words, and then learn to fit the discovered pose-aware transitive words to human pose estimation model through a transfer learning procedure. This paper investigate how to discover and how to use higher-level pose semantics, which can be extended to any pose estimation baselines. Furthermore, we report quantitative and qualitative results on pose analysis with pose retrievals in our built-up data source, and empirical evaluation demonstrates the effectiveness.