An Improved Action Recognition Network With Temporal Extraction and Feature Enhancement

Abstract

Image classification and action recognition are both active research topics in the field of computer vision. However, the development of action recognition is rather slow compared with image classification, due to the difficulties in spatial-temporal information modeling. In this paper, we present TEFE, a deep structure combining temporal extraction with feature enhancement to explore the spatial coherence across temporal dimension. The temporal extraction (TE) module is used to capture the short-term and long-term temporal features. Considering the spatial and temporal cues are fine-grained, we believe such cues (if encoded by well-designed bilinear models) could enhance the representation of actions in videos. The feature enhancement (FE) module approximates bilinear pooling operations, which greatly reduce the amount of parameters exist in original bilinear pooling. Extensive experiments have been conducted on mainstream datasets of human action (Something -Something V1 & V2, and Jester). Experimental results show that our model achieves competitive performances than some existing methods.

Dilated Multi-Temporal Modeling for Action Recognition

“…DMTNet achieved a 1% improvement in performance with 8 frames and a 1.8% improvement with 16 frames compared to TANet, which is our baseline that also uses the same attention strategy and adoptive kernel generating for depth-wise convolution. TEFE [34] showed competitiveness at the 16-frame sampling, but DMTNet exhibited significant results for real-time applications such as for 8-frame rapid inference. The results were obtained by testing the validation set.…”

Sports Teaching Action Recognition Based on Hybrid CNN-HMM

“…Before starting DMA transmission, DMA request should be issued. When a signal meeting the trigger condition enters the data acquisition card, the logic circuit (FPGA or PLD) on the board will drive the ADC to start sampling [6][7][8]. The timing function of the universal timer is enabled during programming.…”

ViGAT: Bottom-Up Event Recognition and Explanation in Video Using Factorized Graph Attention Network

“…In [72], a differentiable similarity guided sampling module is introduced in the architecture of 3D-CNNs that measures the similarity of temporal feature maps and adaptively adjusts the temporal resolution. In [1], an efficient architecture is proposed, consisting of a 2D-CNN and two lightweight 1D-CNN-based branches to capture spatial information, short-and long-term motion dynamics, respectively, and a 3D-CNN feature enhancement module to obtain more fine-grained spatial and temporal cues. This architecture is much more efficient from SlowFast, which uses two 3D-ResNets in its branches.…”

“…, N . The feature representations are stacked row-wise to obtain matrix Γ ∈ R N ×F , Γ = [γ (1) , . .…”

“…Towards this direction, the recognition of high-level events and actions in unconstrained videos plays a crucial role for improving the quality of provided services in various applications, e.g. [1]- [7].…”