Learning contrastive feature distribution model for interaction recognition

• We introduce an intra–inter-frame skeleton feature for interaction description.
• We learn CFDM for a discriminative representation of interactions.
• We capture a new database of interactions, CR-UESTC.
• We evaluate our proposed CFDM approach on CR-UESTC and SBU interaction databases.
• CFDM performs better than CM and BoW, and obtains a higher accuracy than previous works.

In this paper, we learn a Contrastive Feature Distribution Model (CFDM) for interaction recognition. Our contributions are three-folded. First of all, we introduce an intra–inter-frame skeleton feature for interaction description. Secondly, we learn CFDM for a discriminative representation of interactions. In this step, we mine contrastive features to create a dictionary, and learn the probability distribution of dictionary words to construct CFDM in positive and negative training samples. With CFDM, we represent interactions in a discriminative way for recognition. Since there is few skeleton based interaction databases now, we capture a new database, CR-UESTC, which is the third contribution. We evaluate the proposed CFDM approach on CR-UESTC and SBU interaction databases, and compare the result of CFDM with the CM and the BoW approach. The comparison indicates that the recognition accuracy of three approaches is: CFDM > CM > BoW. Compared with Yun et al. (2012), the proposed CFDM also obtain a better result on SBU database.