DG-STA
None
view repo
We propose a Dynamic Graph-Based Spatial-Temporal Attention (DG-STA) method for hand gesture recognition. The key idea is to first construct a fully-connected graph from a hand skeleton, where the node features and edges are then automatically learned via a self-attention mechanism that performs in both spatial and temporal domains. We further propose to leverage the spatial-temporal cues of joint positions to guarantee robust recognition in challenging conditions. In addition, a novel spatial-temporal mask is applied to significantly cut down the computational cost by 99 experiments on benchmarks (DHG-14/28 and SHREC'17) and prove the superior performance of our method compared with the state-of-the-art methods. The source code can be found at https://github.com/yuxiaochen1103/DG-STA.
READ FULL TEXT VIEW PDFNone
Hand gesture recognition has been an active research area due to its wide range of applications such as human computer interaction, gaming and nonverbal communication analysis including sign language recognition [Nikam and Ambekar(2016), Wachs et al.(2011)Wachs, Kölsch, Stern, and Edan, Rautaray and Agrawal(2015)]
. Previous work can be classified into two categories based on the modality of their inputs: image-based
[Freeman and Roth(1995), Molchanov et al.(2015)Molchanov, Gupta, Kim, and Kautz, Wang et al.(2015)Wang, Liu, and Chan] and skeleton-based [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre, Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang, De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat, Chen et al.(2017)Chen, Guo, Wang, and Zhang, Núñez et al.(2018)Núñez, Cabido, Pantrigo, Montemayor, and Vélez] methods. Image-based methods take RGB or RGB-D images as inputs and rely on image-level features for recognition. On the other hand, skeleton-based methods make predictions by a sequence of hand joints with 2D or 3D coordinates. They are more robust to varying lighting conditions and occlusions given the accurate joint coordinates. Thanks to the low-cost depth cameras (e.g, Microsoft Kinect or Intel RealSense) and great progress made on hand pose estimation
[Oberweger et al.(2015a)Oberweger, Wohlhart, and Lepetit, Oberweger and Lepetit(2017), Oberweger et al.(2015b)Oberweger, Wohlhart, and Lepetit], accurate coordinates of hand joints are easy to be obtained. Therefore, we follow the skeleton-based method in this work.Conventional methods [Ohn-Bar and Trivedi(2013), De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre, De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat] of skeleton-based hand gesture recognition aim to design powerful feature descriptors to model the action of hands. However, these hand-crafted features have limited generalization capability. Recent studies [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang, Chen et al.(2017)Chen, Guo, Wang, and Zhang, Núñez et al.(2018)Núñez, Cabido, Pantrigo, Montemayor, and Vélez]
have achieved significant improvement using deep learning. They usually concatenate the joint coordinates into a tensor which is fed into a neural network, and then hand features are directly learned by the network during training. Nevertheless, the spatial structures and temporal dynamics of hand skeletons are not explicitly exploited in these deep learning based approaches.
More recent studies [Yan et al.(2018)Yan, Xiong, and Lin, Si et al.(2018)Si, Jing, Wang, Wang, and Tan, Zhao et al.(2019)Zhao, Peng, Tian, Kapadia, and Metaxas]
attempt to incorporate structures and dynamics of skeletons based on skeleton graphs. Specifically, given a sequence of skeletons, they define a spatial-temporal graph where structures and dynamics of skeletons are embedded. The feature representation of the graph is then extracted for action recognition. However, a pre-defined graph with fixed structure lacks the flexibility to capture the variance and dynamics across different actions, yielding sub-optimal performance in practice.
To this end, we propose a Dynamic Graph-Based Spatial Temporal Attention (DG-STA) model for hand gesture recognition. The key idea is to perform a self-attention mechanism in both spatial and temporal domains to modify a unified graph dynamically in order to model different actions. Figure 1 gives an overview of our approach. There are three crucial designs that distinguish our approach from previous methods. First, instead of a pre-defined graph with fixed structure, we propose to construct a unified graph where the edges and nodes are dynamically optimized according to different actions. This makes it is possible to achieve action-specific graphs with improved expressive power. Second, we propose spatial-temporal position embedding which improves the temporal position embedding [Vaswani et al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin]. It encodes the identity and temporal order information of each node in the graph. Combining node features with their position embeddings can further improve the performance of our approach. Third, to implement our DG-STA more efficiently, we present a novel spatial-temporal mask operation which is directly applied to the matrix of scaled dot-products among all nodes. It significantly improves the computational efficiency of our model and allows easier input data arrangement.
To evaluate the effectiveness of our approach, we conduct comprehensive experiments on two standard benchmarks: DHG-14/28 Dataset [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre] and SHREC’17 Track Dataset [De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat]. The results demonstrate that our method outperforms the state-of-the-art methods. In summary, our main contributions are summarised as follows:
We propose Dynamic Graph-Based Spatial-Temporal Attention (DG-STA) for skeleton-based hand gesture recognition. The structures and dynamics of hand skeletons can be learned automatically and more efficiently by our approach.
We propose spatial-temporal position embedding which encodes the identity and temporal order information of nodes to boost the performance of our model, and a spatial-temporal mask operation for efficient implementation of DG-STA.
We conduct comprehensive experiments to validate our approach on two standard benchmarks. The proposed DG-STA achieves the state-of-the-art performance.
In this section, we review recent work on self-attention and recent developments in skeleton-based action recognition, which motivated our approach.
Self-Attention.
The self-attention mechanism is widely used in computer vision and natural language processing tasks
[Chen et al.(2018b)Chen, Zhang, You, Fang, Wang, Jin, and Luo, Chen et al.(2018c)Chen, Yuan, You, and Luo, Chen et al.(2018a)Chen, Chen, Guo, and Luo, Zhang et al.(2019)Zhang, Goodfellow, Metaxas, and Odena, Lin et al.(2017)Lin, Feng, Santos, Yu, Xiang, Zhou, and Bengio, Tan et al.(2018)Tan, Wang, Xie, Chen, and Shi, Verga et al.(2018)Verga, Strubell, and McCallum, Tian et al.(2018)Tian, Peng, Zhao, Zhang, and Metaxas]. Vaswani et al [Vaswani et al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin] proposed to apply the self-attention module to model temporal and semantic relationships among words within a sentence for machine translation. Instead, we study applying the self-attention mechanism to learn spatial-temporal information contained in hand skeletons represented by graphs, which are largely different from sequences. Graph Attention Networks (GATs) [Veličković et al.(2017)Veličković, Cucurull, Casanova, Romero, Lio, and Bengio] employed self-attention to learn node embeddings of graphs. By contrast, our approach is able to capture additional temporal information as well as node identities.Skeleton-Based Hand Gesture Recognition. Skeleton-based hand gesture is a well-studied but still challenging task. Traditional methods [Ohn-Bar and Trivedi(2013), De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre, De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat, Lu et al.(2003)Lu, Metaxas, Samaras, and Oliensis] mainly focus on designing powerful hand feature descriptors. Smedt et al [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre] proposed the Shape of Connected Joints descriptor to represent the hand shape of hand skeleton. Recent studies [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang, Chen et al.(2017)Chen, Guo, Wang, and Zhang, Núñez et al.(2018)Núñez, Cabido, Pantrigo, Montemayor, and Vélez]
apply deep neural networks for this task and achieve significant performance improvement. Convolutional Neural Networks and Long Short-Term Memory are leveraged to learn the spatial and temporal features from the sequence of hand joints for hand gesture classification in
[Núñez et al.(2018)Núñez, Cabido, Pantrigo, Montemayor, and Vélez]. One limitation of these learning-based methods is that they do not explicitly explore the structures and dynamics of human hands.Skeleton-Based Human Action Recognition. Recent studies in skeleton-base human action recognition [Zhao et al.(2018)Zhao, Peng, Tian, Kapadia, and Metaxas, Tang et al.(2018)Tang, Peng, Geng, Wu, Zhang, and Metaxas, Peng et al.(2018)Peng, Tang, Yang, Feris, and Metaxas] started to incorporate structures and dynamics of human bodies by building skeleton graphs [Yan et al.(2018)Yan, Xiong, and Lin, Si et al.(2018)Si, Jing, Wang, Wang, and Tan, Zhao et al.(2019)Zhao, Peng, Tian, Kapadia, and Metaxas]. This idea is first introduced by [Edwards and Xie(2017)] which employs Graph CNNs. Recently, Yan et al [Yan et al.(2018)Yan, Xiong, and Lin] built a skeleton graph based on the natural structure of human body, and extract its representation by Graph Convolution Networks [Kipf and Welling(2016)] for action recognition. Nevertheless, it is difficult to define an optimal skeleton graph which represents all action-specific structures and dynamics information. Instead, our method can automatically learn multiple action-specific graphs with the multi-head attention mechanism [Vaswani et al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin], which efficiently encode structures and dynamics of hand gestures.
The overview of our approach is shown in Figure 1. First, a fully-connected skeleton graph is constructed from the input sequence of hand skeletons as described in Section 3.1. In Section 3.2
, we devise DG-STA to learn the edge weights and node embeddings within the graph. The learned node features by DG-STA are then average-pooled into a vector which captures the structures and dynamics of the input skeleton graph. We use it for hand gesture classification. Section
3.3 presents the spatial-temporal position embedding which is combined with node features to incorporate node identity and temporal order information contained in the hand skeletons. Moreover, a spatial-temporal mask operation is introduced in Section 3.4 which implements our proposed DG-STA more efficiently.Given a video of frames, hand joints are extracted from each frame to represent the hand skeleton. Then a fully-connected skeleton graph is constructed from this sequence of hand skeletons. Let denote the node set where represents the -th hand joint at the time step . The node features are represented by , where indicates the feature vector of the node . They are extracted from the 3D coordinates of nodes. Note that each node is connected with all other nodes including itself. For clarity, we define three types of edges on the edge set as follows.
A spatial edge connects two different nodes at the same time step.
A temporal edge connects two nodes at different time steps.
A self-connected edge connects the node with itself.
The proposed DG-STA consists of the spatial attention model
and temporal attention model which are employed to extract spatial and temporal information from the hand skeleton graph respectively. Both and are based on multi-head attention [Vaswani et al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin]. first takes the initial node features as the input and updates them to encode spatial information. The updated node features are then fed to to further learn temporal information. Finally, the results are average-pooled to a vector which is used as the feature representation of the skeleton graph for classification.Specifically, given the input feature of the node in the skeleton graph, the -th spatial attention head first applies three fully-connected layers to map into the key, query and value vectors respectively, which are formulated as:
(1) |
where , and represent the key, query and value vectors of the node; , and are the corresponding weight matrices of the three fully-connected layers of the -th spatial attention head.
The spatial attention head computes the weights of the spatial and self-connected edges in two steps. First, it calculates the “scaled dot-product” [Vaswani et al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin] between the query vectors and key vectors of the nodes within the same time step. Then it normalizes the results by a Softmax function. These two steps are formulated as:
(2) |
where is the dimension of the key, query and value vectors; is the scaled dot-product of the node and ; represents the inner product operation; is the attention weight between the node and , which measures the importance of information from node to node . Meanwhile, the weights for all temporal edges are set to 0 in order to block the information passing in the temporal domain. As a result, each spatial attention head produces a weighted skeleton graph which represents a specific type of spatial structure of the hand.
The attention head calculates the spatial attention feature of the node as the weighted sum of the value vectors within the same time step, which is defined as:
(3) |
where denotes the spatial attention feature of the node . Intuitively, the computation mechanism of the spatial attention features is essentially the process that each node in the graph sends some information to the others within the same time step and then aggregates the received information based on the learned edge weights.
The spatial attention model finally concatenates the spatial attention features learned by all spatial attention heads into which is employed as the spatial feature of the node :
(4) |
where is the number of spatial attention heads. The obtained node features encode multiple types of structural information represented by the weighted skeleton graphs which are learned by different spatial attention heads.
The temporal attention model takes the output node features from the spatial attention model as the input, and then applies the above multi-head attention mechanism in the temporal domain. The node feature which is the output from the temporal attention model encodes both spatial and temporal information carried by the input sequence of the hand skeletons. We average-pool these node features to a vector as the feature representation of the input sequence for hand gesture recognition.
The original node features extracted from the coordinates of the input hand skeletons do not contain spatial identity information describing which hand joint a node corresponds to, and temporal information indicating which time step a node is at. To incorporate these messages, we propose the spatial-temporal position embedding.
Specifically, our spatial-temporal position embedding is made up of the spatial position embedding and the temporal position embedding. The spatial one consists of vectors and each represents a hand joint. Meanwhile, the temporal one is composed of distinct vectors and each of them corresponds to a node in the hand skeleton graph. The feature vector of a specific node is added with the corresponding spatial and temporal position embedding vectors before fed into the DG-STA. Therefore, we have:
(5) |
where denotes the final output feature of node , is the spatial position embedding of the -th hand joint, and denotes the temporal position embedding of the -th hand joint at -th time step. These embeddings are with the same dimension as , and their values are set using the sine and cosine functions of different frequencies following [Vaswani et al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin].
It is not straightforward to implement the proposed DG-STA because the input data have to be arranged in a complex format. However, we find that the computation of attention weights and features without domain constraints is straightforward, which can be implemented efficiently using matrix multiplication operations. Therefore, we propose a novel scheme to facilitate the implementation of the DG-STA. The main idea is to first compute the matrix of the scaled dot-products among all nodes and then apply the proposed spatial-temporal mask operation to the matrix in order to let the model focus on the spatial or temporal domain.
An illustration of our mask operation is shown in Figure 2. For a specific attention head, we compute a query matrix where each row represents the query vector of each node, and a key matrix where each row corresponds to the key vector of each node. The matrix of the scaled dot-products (i.e, the edge weights before normalization) can be obtained by:
(6) |
where is the matrix multiplication, and denotes the matrix transpose operation. Then the proposed spatial mask operation sets the value of each element in which represents the temporal edge to (i.e, a number close to negative infinity) and keeps the values of other elements unchanged. Therefore, the resulting matrix after the spatial mask operation is calculated:
(7) |
where denotes the element-wise dot operation, is the spatial mask where the elements are 1 if they represent the spatial or self-connect edges and 0 otherwise. We set to in our implementation. The Softmax activation essentially normalizes weights across spatial edges, because the exponential value of the is close to 0. As a result, all weights of the temporal edges in are set to 0. Equations (6) and (7) efficiently implement the calculation of edge weights in the spatial domain formulated in Equation (2). Moreover, the matrix can be directly employed to implement the computation of node features represented by Equation (3) by performing the matrix multiplication operation with the matrix of value vectors.
We define the temporal mask operation following the same way as Equation (7). The difference is that we use the temporal mask instead of to compute the matrix after the temporal mask operation . The elements of are 1 if they represent the temporal or self-connect edges and 0 otherwise. With the help of the proposed spatial-temporal mask operation, we experimentally find that the computation time is reduced by 99%.
In this section, we first describe our network structure in Section 4.1. In Section 4.2, we introduce the datasets and settings employed in the experiments. Then we conduct ablation studies in Section 4.3 to evaluate the effectiveness of each component proposed in our method. Finally, we report our results and comparisons with the state of the art in Section 4.4.
Our network structure is shown in Figure 3. We set the head number of the spatial and temporal attention models to 8. The dimension of the query, key and value vector is set to 32. Layer Normalization [Lei Ba et al.(2016)Lei Ba, Kiros, and Hinton] is utilized to normalize the intermediate outputs of our network. The input 3D coordinate of a hand joint is projected into an initial node feature of 128 dimension. It is then added with the corresponding spatial position embedding and fed into the spatial attention model, which produces a node feature of 256 dimension. This node feature is projected into a vector of 128 dimension which is added with the corresponding temporal position embedding. The temporal attention model takes it as the input and generates the final node feature. Finally, we average-pool the features of all nodes into a vector and feed it into a fully-connected layer for classification.
We evaluate our method on the DHG-14/28 Dataset [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre] and the SHREC’17 Track Dataset [De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat]. Both datasets contain 2800 video sequences of 14 hand gestures which are performed in two configurations: using one single finger or the whole hand. The videos of the two datasets are captured by the Intel Realsense camera. The 3D coordinates of 22 hand joints in real-world space are provided per frame for network training and evaluation.
Network Training.
The proposed DG-STA is implemented based on the PyTorch platform. The Adam
[Kingma and Ba(2014)] optimizer with a learning rate of 0.001 is employed to train our model. The batch size is set to 32 and the dropout rate [Srivastava et al.(2014)Srivastava, Hinton, Krizhevsky, Sutskever, and Salakhutdinov] is set to 0.2. We uniformly sample 8 frames from each video as the input. For fair comparison, we perform data augmentation by applying the same operations as proposed in [De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat, Núñez et al.(2018)Núñez, Cabido, Pantrigo, Montemayor, and Vélez]including scaling, shifting, time interpolation and adding noise. We also subtract every skeleton sequence by the palm position of the first frame following Smedt
et al [De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat] for alignment.Evaluation Protocols. On the DHG-14/28 Dataset, models are evaluated by using the leave-one-subject-out cross-validation strategy [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre]. Specifically, we perform one experiment for each subject in this dataset. In each experiment, one subject is selected for testing and the remaining 19 subjects are used for training. The average accuracy of 14 gestures (without the single-finger configuration) or 28 gestures (with the single-finger configuration) over the 20 cross-validation folds are reported. For the SHREC’17 Track Dataset, we use the same data split as provided by [De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat] and report the accuracy of both 14 and 28 gestures.
Our proposed approach consists of three major components, including the fully-connected skeleton graph structure (FSG), the spatial-temporal attention model (STA) and the spatial-temporal position embedding (STE). We validate the effectiveness of these components in this section. The results are shown in Table 1.
Setting | FSG+STA | FSG+GAT+STE | SSG+STA+STE | DG-STA |
---|---|---|---|---|
14 Gestures (D) | 84.3 | 90.8 | 89.8 | 91.9 |
28 Gestures (D) | 77.3 | 87.8 | 86.6 | 88.0 |
14 Gestures (S) | 88.9 | 92.7 | 91.5 | 94.4 |
28 Gestures (S) | 80.1 | 86.2 | 87.7 | 90.7 |
Evaluation of Fully-Connected Graph Structure. We compare the proposed FSG with the sparse skeleton graph structure (SSG) introduced by Yan et al [Yan et al.(2018)Yan, Xiong, and Lin], where spatial edges are defined based on the natural connections of the hand joints and temporal edges connect the same joints between consecutive frames. We can see that our model significantly outperforms the one trained on SSG. This is because SSG may be sub-optimal for some hand gestures, while FSG has little constrains on the model so that it is able to learn action-specific graph structures.
Evaluation of Spatial-Temporal Attention. The proposed STA downgrades to Graph Attention (GAT) [Veličković et al.(2017)Veličković, Cucurull, Casanova, Romero, Lio, and Bengio] if only one attention model is applied to the whole graph without distinguishing the spatial and temporal domains. We implement GAT by replacing the spatial and temporal attention models in our network with one attention model, and train it under the same setting of our model. We can observe that the STA-based model achieves better performance than the GAT-based model, which demonstrates the effectiveness of STA.
Evaluation of Spatial-Temporal Position Embedding. We validate the effectiveness of the proposed STE by training a variant of our method where STE is removed. We can see that our model outperforms the model without STE, which demonstrates the importance of the identity and temporal order information encoded by STE.
We compare our method with the state-of-the-art methods on the DHG-14/28 Dataset [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre] and the SHREC’17 Track Dataset [De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat], respectively. The compared state-of-the-art methods include traditional hand-crafted feature approaches [Chen et al.(2017)Chen, Guo, Wang, and Zhang, Oreifej and Liu(2013), Devanne et al.(2015)Devanne, Wannous, Berretti, Pala, Daoudi, and Del Bimbo, Ohn-Bar and Trivedi(2013), De Smedt(2017), De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre, Caputo et al.(2018)Caputo, Prebianca, Carcangiu, Spano, and Giachetti, Boulahia et al.(2017)Boulahia, Anquetil, Multon, and Kulpa], deep learning based approaches [Núñez et al.(2018)Núñez, Cabido, Pantrigo, Montemayor, and Vélez, Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang, De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat] and a graph-based method [Yan et al.(2018)Yan, Xiong, and Lin]. The results are shown in Tables 2 and 3. Note that for ST-GCN [Yan et al.(2018)Yan, Xiong, and Lin], we implement it following the distance partitioning setting and use a three-layer ST-GCN with 128 channels for fair comparison. We collect the results of other baseline methods from [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang].
Method | 14 Gestures | 28 Gestures |
---|---|---|
SoCJ+HoHD+HoWR [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre] | 83.1 | 80.0 |
Chen et al [Chen et al.(2017)Chen, Guo, Wang, and Zhang] | 84.7 | 80.3 |
CNN+LSTM [Núñez et al.(2018)Núñez, Cabido, Pantrigo, Montemayor, and Vélez] | 85.6 | 81.1 |
Res-TCN [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang] | 86.9 | 83.6 |
STA-Res-TCN [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang] | 89.2 | 85.0 |
ST-GCN [Yan et al.(2018)Yan, Xiong, and Lin] | 91.2 | 87.1 |
DG-STA (Ours) | 91.9 | 88.0 |
Results on DHG-14/28 Dataset. From Table 2, we can see that our method achieves the state-of-the-arts performance under both 14-gesture and 28-gesture setting. Moreover, both our method and ST-GCN [Yan et al.(2018)Yan, Xiong, and Lin] outperform other methods which do not explicitly exploit structures and dynamics of hands, which demonstrates that these messages are important for skeleton-based hand gesture recognition.
Method | 14 Gestures | 28 Gestures |
---|---|---|
Oreifej et al [Oreifej and Liu(2013)] | 78.5 | 74.0 |
Devanne et al [Devanne et al.(2015)Devanne, Wannous, Berretti, Pala, Daoudi, and Del Bimbo] | 79.4 | 62.0 |
Classify Sequence by Key Frames [De Smedt et al.(2017)De Smedt, Wannous, Vandeborre, Guerry, Le Saux, and Filliat] | 82.9 | 71.9 |
Ohn-Bar et al [Ohn-Bar and Trivedi(2013)] | 83.9 | 76.5 |
SoCJ+Direction+Rotation [De Smedt(2017)] | 86.9 | 84.2 |
SoCJ+HoHD+HoWR [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre] | 88.2 | 81.9 |
Caputo et al [Caputo et al.(2018)Caputo, Prebianca, Carcangiu, Spano, and Giachetti] | 89.5 | - |
Boulahia et al [Boulahia et al.(2017)Boulahia, Anquetil, Multon, and Kulpa] | 90.5 | 80.5 |
Res-TCN [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang] | 91.1 | 87.3 |
STA-Res-TCN [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang] | 93.6 | 90.7 |
ST-GCN [Yan et al.(2018)Yan, Xiong, and Lin] | 92.7 | 87.7 |
DG-STA (Ours) | 94.4 | 90.7 |
Results on SHREC’17 Track Dataset. Different from the DHG-14/28 Dataset where videos are cropped by human-labeled beginnings and ends of the gestures [De Smedt et al.(2016)De Smedt, Wannous, and Vandeborre], the SHREC’17 Track Dataset provides raw captured video sequences with noisy frames, and hence is more challenging. We can see that our method achieves the state-of-the-arts performance under the 14-gesture setting, and obtains comparable performance with STA-Res-TCN [Hou et al.(2018)Hou, Wang, Chen, Xue, Zhu, and Yang] under the 28-gesture setting. In addition, we can observe that our method and ST-GCN [Yan et al.(2018)Yan, Xiong, and Lin] outperform all other methods which do not explicitly exploit structures and dynamics of hands.
In this paper, we proposed a graph-based spatial-temporal attention method for skeleton-based hand-gesture recognition. It utilizes two attention models in the spatial and temporal domains of the fully-connected hand skeleton graph to learn edge weights and extract spatial and temporal information for hand gesture recognition. Extensive experiments demonstrate the effectiveness of our framework. Our proposed method provides a general framework that can be further used for other tasks aiming to learn spatial and temporal information from graph-based data, e.g, skeleton-based human action recognition.
This work was funded partly by ARO-MURI-68985NSMUR and NSF 1763523, 1747778, 1733843, 1703883 to Dimitris N. Metaxas.
When e-commerce meets social media: Identifying business on wechat moment using bilateral-attention lstm.
In Proceedings of the World Wide Web Conference (WWW), pages 343–350, 2018a.Motion feature augmented recurrent neural network for skeleton-based dynamic hand gesture recognition.
In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 2881–2885, 2017.Twitter sentiment analysis via bi-sense emoji embedding and attention-based lstm.
In Proceedings of the ACM Multimedia Conference on Multimedia Conference (MM), pages 117–125, 2018c.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)
, pages 1–9, 2016.The Journal of Machine Learning Research
, 15(1):1929–1958, 2014.