On the retrieval of 3D mesh sequences of human actions

Unsupervised human action retrieval using salient points in 3D mesh sequences

“…9. The retrieval results of the other methods are related to the optimal values presented in [27]. Again, the proposed method almost always performs best and the reduction in retrieval performance is low compared to the corresponding results using the full sequences of the same dataset.…”

“…8. The retrieval results of the other methods are related to the optimal values presented in [27]. As can be seen, the proposed method achieves ideal, or in some metrics almost ideal, results, outperforming competing methods.…”

“…Directly comparable to the proposed approach, is the work in [27]. In this work, a set of state-of-the-art static shape descriptors are compared in human action retrieval framework.…”

“…The retrieval results of other methods are related to the optimal values presented in [27]. [27], [17] 0.733 0.547 0.703 0.761 PANORAMA [27], [18] 0.717 0.553 0.650 0.748 Shape Distribution [27], [16] 0.633 0.367 0.563 0.638 Spin Images [27], [9] 0 Fig. 7 Precision-recall diagrams for the DUTH dataset.…”

“…As can be seen, the proposed method achieves ideal, or in some metrics almost ideal, results, outperforming competing methods. [27], [17] 0.980 0.968 0.999 0.994 PANORAMA [27], [18] 0.985 0.973 1.000 0.992 Shape Distribution [27], [16] 0.921 0.889 0.972 0.956 Spin Images [27], [9] 1.000 0.871 0.941 0.972 Proposed Method 1.000 0.998 1.000 1.000 In order to cancel the 1-1 correspondence between the frames of the actions of the USurrey dataset, an additional experiment has been performed. Specifically, corresponding sequences have been truncated by a random number of frames (up to 50) from the starting frame.…”