Journal
PATTERN RECOGNITION
Volume 103, Issue -, Pages -Publisher
ELSEVIER SCI LTD
DOI: 10.1016/j.patcog.2020.107321
Keywords
Graph convolutional network; Structure graph pooling; Joint-wise channel attention
Funding
- National Key RD Plan [2017YFB1-002801, 2016QY01W0106]
- Natural Science Foundation of China [U1803119, U1736106, 61751212, 61721004, 61972397, 61772225]
- NSFC-General Technology Collaborative Fund for Basic Research [U1636218]
- Key Research Program of Frontier Sciences, CAS [YZDJSSW-JSC040]
- Beijing Natural Science Foundation [JQ18018, L172051, L182058]
- CAS External Cooperation Key Project
- Youth Innovation Promotion Association, CAS
Ask authors/readers for more resources
Recently, graph convolutional networks (GCNs) have achieved state-of-the-art results for skeleton based action recognition by expanding convolutional neural networks (CNNs) to graphs. However, due to the lack of effective feature aggregation method, e.g. max pooling in CNN, existing GCN-based methods only learn local information among adjacent joints and are hard to obtain high-level interaction features, such as interactions between five parts of human body. Moreover, subtle differences of confusing actions often hide in specific channels of key joints' features, this kind of discriminative information is rarely exploited in previous methods. In this paper, we propose a novel graph convolutional network with structure based graph pooling (SGP) scheme and joint-wise channel attention UCA) modules. The SGP scheme pools the human skeleton graph according to the prior knowledge of human body's typology. This pooling scheme not only leads to more global representations but also reduces the amount of parameters and computation cost. The JCA module learns to selectively focus on discriminative joints of skeleton and pays different levels of attention to different channels. This novel attention mechanism enhance the model's ability to classify confusing actions. We evaluate our SGP scheme and JCA module on three most challenging skeleton based action recognition datasets: NTU-RGB+D, Kinetics-M, and SYSU-3D. Our method outperforms the state-of-art methods on three benchmarks. (C) 2020 Elsevier Ltd. All rights reserved.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available