Human Activities as Stochastic Kronecker Graphs

Abstract

Abstract.A human activity can be viewed as a space-time repetition of activity primitives. Both instances of the primitives, and their repetition are stochastic. They can be modeled by a generative model-graph, where nodes correspond to the primitives, and the graph's adjacency matrix encodes their affinities for probabilistic grouping into observable video features. When a video of the activity is represented by a graph capturing the space-time layout of video features, such a video graph can be viewed as probabilistically sampled from the activity's modelgraph. This sampling is formulated as a successive Kronecker multiplication of the model's affinity matrix. The resulting Kronecker-power matrix is taken as a noisy permutation of the adjacency matrix of the video graph. The paper presents our: 1) model-graph; 2) memory-and time-efficient, weakly supervised learning of activity primitives and their affinities; and 3) inference aimed at finding the best expected correspondences between the primitives and observed video features. Our results demonstrate good scalability on UCF50, and superior performance to that of the state of the art on individual, structured, and collective activities of UCF YouTube, Olympic, and Collective datasets.

Regression-Based Feature Selection on Large Scale Human Activity Recognition

“…The comparisons to available related works are described in Table II. www.ijacsa.thesai.org [14] group-wise cross validation 57.90% Todorovic [23] 2/3 training and 1/3 testing for each class 81.03% Solmaz et al [24] Leave One Group Out Cross validation(25 cross-validations) 73.70%…”

A Time Flexible Kernel framework for video-based activity recognition

“…Thanks to these restrictions it is possible to train states encompassing common characteristics among the videos and thus to achieve high accuracy. However, new databases, such as HMDB51, UCF50, OlympicSports and Virat Release 2.0 have been produced aiming at challenging tasks, such as indexing events in 35 unconstrained videos in the internet or surveillance in uncontrolled scenarios performing a scene-independent learning and recognition. These datasets were recorded in unconstrained environments with random viewpoints, camera movements and/or dynamic changes in the background.…”

Robust Pose Features for Action Recognition

“…Le et al [10] learn features directly from video using independent subspace analysis that is robust to translation and selective to frequency and rotation changes. Todorovic [17] views a human activity as a space-time repetition of activity primitives and models the primitives and their repetition by a generative model-graph. Sadanand and Corso [16] propose action bank, consisting of action detectors sampled according to classes and viewpoints.…”

“…For each feature, we set the initial vocabulary size to 500 and the semantic vocabulary size to 100. During training, we store for each poselet the video Method Accuracy (%) Wang et al [18] 85.6 Le et al [10] 86.5 O'Hara and Draper [14] 91.3 Todorovic [17] 92.1 Sadanand and Corso [16] 95 sequence from which its training images were selected. For clustering, we set the portion of coverage to 0.8, resulting in 123, 120, 120, and 123 poselets for the root and the three parts, respectively.…”