An ensemble approach for still image-based human action recognition

Still-image based human action recognition is a challenging task in the field of computer vision due to the limited information available in a single image. Hence, it is important to efficiently extract visual cues and structural information from the image in the process of classification. To this end, in this work, we utilize the Convolutional neural network for classification, based on the DenseNet 201 architecture. To focus upon informative regions of interest, the spatial attention module has been trained as a feature extractor to emphasize features from selective parts of the input image. We further leverage an effective ensemble approach based upon fuzzy fusion through the Choquet integral, which harnesses the complementary uncertainty of decision scores. This allows for a robust decision-making process on the fly, based upon coalitions of the inputs. Experimental results upon three challenging datasets: PPMI and Stanford 40, known for their confusing action classes, and BU-101, known for its immense scale, support the efficacy of the proposed method.