Visual attention has been shown effective in various structural prediction tasks such as image/video captioning [34, 36] and visual question answering [4, 35, 33]. Its success is mainly due to the reasonable assumption that human vision does not tend to process a whole image in its entirety at once; instead, one only focuses on selective parts of the whole visual space when and where as needed 
. Specifically, rather than encoding an image into a static vector, attention allows the image feature to evolve from the sentence context at hand, resulting in richer and longer descriptions for cluttered images. In this way, visual attention can be considered as a dynamic feature extraction mechanism that combines contextual fixations over time[19, 26].
State-of-the-art image features are generally extracted by deep Convolutional Neural Networks (CNNs) [8, 25, 32]. Starting from an input color image of the size , a convolutional layer consisting of -channel filters scans the input image and output a feature map, which will be the input for the next convolutional layer111Each convolutional layer is optionally followed by a pooling, down-sampling, normalization, or a fully connected layer.. Each 2D slice of a 3D feature map encodes the spatial visual responses raised by a filter channel, where the filter performs as a pattern detector — lower-layer filters detect low-level visual cues like edges and corners while higher-level ones detect high-level semantic patterns like parts and object . By stacking the layers, a CNN extracts image features through a hierarchy of visual abstractions. Therefore, CNN image features are essentially spatial, channel-wise, and multi-layer. However, most existing attention-based image captioning models only take into account the spatial characteristic , i.e., those attention models merely modulate the sentence context into the last conv-layer feature map via spatially attentive weights.
In this paper, we will take full advantage of the three characteristics of CNN features for visual attention-based image captioning. In particular, we propose a novel Spatial and Channel-wise Attention-based Convolutional Neural Network, dubbed SCA-CNN, which learns to pay attention to every feature entry in the multi-layer 3D feature maps. Figure 1
illustrates the motivation of introducing channel-wise attention in multi-layer feature maps. First, since a channel-wise feature map is essentially a detector response map of the corresponding filter, channel-wise attention can be viewed as the process of selecting semantic attributes on the demand of the sentence context. For example, when we want to predictcake, our channel-wise attention (e.g., in the / feature map) will assign more weights on channel-wise feature maps generated by filters according to the semantics like cake, fire, light, and candle-like shapes. Second, as a feature map is dependent on its lower-layer ones, it is natural to apply attention in multiple layers, so as to gain visual attention on multiple semantic abstractions. For example, it is beneficial to emphasize on lower-layer channels corresponding to more elemental shapes like array and cylinder that compose cake.
We validate the effectiveness of the proposed SCA-CNN on three well-known image captioning benchmarks: Flickr8K, Flickr30K and MSCOCO. SCA-CNN can significantly surpass the spatial attention model  by in BLEU4. In summary, we propose a unified SCA-CNN framework to effectively integrate spatial, channel-wise, and multi-layer visual attention in CNN features for image captioning. In particular, a novel spatial and channel-wise attention model is proposed. This model is generic and thus can be applied to any layer in any CNN architecture such as popular VGG  and ResNet . SCA-CNN helps us gain a better understanding of how CNN features evolve in the process of the sentence generation.
2 Related Work
We are interested in visual attention models used in the encoder-decoder framework for neural image/video captioning (NIC) and visual question answering (VQA), which fall into the recent trend of connecting computer vision and natural language[14, 41, 24, 23, 42, 12]. Pioneering work on NIC [31, 13, 6, 30, 29] and VQA [1, 17, 7, 21] uses a CNN to encode an image or video into a static visual feature vector and then feed it into an RNN  to decode language sequences such as captions or answers.
However, the static vector does not allow the image feature adapting to the sentence context at hand. Inspired by the attention mechanism introduced in machine translation , where a decoder dynamically selects useful source language words or sub-sequence for the translation into a target language, visual attention models have been widely-used in NIC and VQA. We categorize these attention-based models into the following three domains that motivate our SCA-CNN:
Spatial Attention. Xu et al.  proposed the first visual attention model in image captioning. In general, they used “hard” pooling that selects the most probably attentive region, or “soft” pooling that averages the spatial features with attentive weights. As for VQA, Zhu et al.  adopted the “soft” attention to merge image region features. To further refine the spatial attention, Yang et al.  and Xu et al.  applied a stacked spatial attention model, where the second attention is based on the attentive feature map modulated by the first one. Different from theirs, our multi-layer attention is applied on the multiple layers of a CNN. A common defect of the above spatial models is that they generally resort to weighted pooling on the attentive feature map. Thus, spatial information will be lost inevitably. More seriously, their attention is only applied in the last conv-layer, where the size of receptive field will be quite large and the differences between each receptive field region are quite limited, resulting in insignificant spatial attentions.
Semantic Attention. Besides the spatial information, You et al. 
proposed to select semantic concepts in NIC, where the image feature is a vector of confidences of attribute classifiers. Jiaet al.  exploited the correlation between images and their captions as the global semantic information to guide the LSTM generating sentences. However, these models require external resources to train these semantic attributes. In SCA-CNN, each filter kernel of a convolutional layer servers as a semantic detectors . Therefore, the channel-wise attention of SCA-CNN is similar to semantic attention.
Multi-layer Attention. According to the nature of CNN architecture, the sizes of respective fields corresponding to different feature map layers are different. To overcome the weakness of large respective field size in the last conv-layer attention, Seo et al.  proposed a multi-layer attention networks. In compared with theirs, SCA-CNN also incorporates the channel-wise attention at multiple layers.
3 Spatial and Channel-wise Attention CNN
We adopt the popular encoder-decoder framework for image caption generation, where a CNN first encodes an input image into a vector and then an LSTM decodes the vector into a sequence of words. As illustrated in Figure 2, SCA-CNN makes the original CNN multi-layer feature maps adaptive to the sentence context through channel-wise attention and spatial attention at multiple layers.
Formally, suppose that we want to generate the -th word of the image caption. At hand, we have the last sentence context encoded in the LSTM memory , where is the hidden state dimension. At the -th layer, the spatial and channel-wise attention weights are a function of and the current CNN features . Thus, SCA-CNN modulates using the attention weights in a recurrent and multi-layer fashion as:
where is the modulated feature, is the spatial and channel-wise attention function that will be detailed in Section 3.2 and 3.3, is the feature map output from previous conv-layer, e.g., convolution followed by pooling, down-sampling or convolution [25, 8], and is a linear weighting function that modulates CNN features and attention weights. Different from existing popular modulating strategy that sums up all visual features based on attention weights , function applies element-wise multiplication. So far, we are ready to generate the -th word by:
where is the total number of conv-layers; is a probability vector and is a predefined dictionary including all caption words.
Note that is of the same size as or , i.e., . It will require space for attention computation, where is the common mapping space dimension of CNN feature and hidden state . It is prohibitively expensive for GPU memory when the feature map size is so large. Therefore, we propose an approximation that learns spatial attention weights and channel-wise attention weights separately:
Where and represent channel-wise and spatial attention model respectively. This will greatly reduce the memory cost into for spatial attention and for channel-wise attention, respectively.
3.2 Spatial Attention
In general, a caption word only relates to partial regions of an image. For example, in Figure 1, when we want to predict cake, only image regions which contain cake are useful. Therefore, applying a global image feature vector to generate caption may lead to sub-optimal results due to the irrelevant regions. Instead of considering each image region equally, spatial attention mechanism attempts to pay more attention to the semantic-related regions. Without loss of generality, we discard the layer-wise superscript . We reshape by flattening the width and height of the original , where and . We can consider as the visual feature of the -th location. Given the previous time step LSTM hidden state , we use a single-layer neural network followed by a softmax function to generate the attention distributions over the image regions. Below are the definitions of the spatial attention model :
where are transformation matrices that map image visual features and hidden state to a same dimension. We denote as the addition of a matrix and a vector. And the addition between a matrix and a vector is performed by adding each column of the matrix by the vector. are model biases.
3.3 Channel-wise Attention
Note that the spatial attention function in Eq (3) still requires the visual feature to calculate the spatial attention weights, but the visual feature used in spatial attention is in fact not attention-based. Hence, we introduce a channel-wise attention mechanism to attend the features . It is worth noting that each CNN filter performs as a pattern detector, and each channel of a feature map in CNN is a response activation of the corresponding convolutional filter. Therefore, applying an attention mechanism in channel-wise manner can be viewed as a process of selecting semantic attributes.
For channel-wise attention, we first reshape to , and , where represents the -th channel of the feature map , and is the total number of channels. Then, we apply mean pooling for each channel to obtain the channel feature :
where scalar is the mean of vector , which represents the -th channel features. Following the definition of the spatial attention model, the channel-wise attention model can be defined as follows:
where are transformation matrices, represents the outer product of vectors. are bias terms.
According to different implementation order of channel-wise attention and spatial attention, there exists two types of model which incorporating both two attention mechanisms. We distinguish between the two types as follows:
Channel-Spatial. The first type dubbed Channel-Spatial (C-S) applies channel-wise attention before spatial attention. The flow chart of C-S type is illustrated in Figure 2. At first, given an initial feature map , we adopt channel-wise attention to obtain the channel-wise attention weights . Through a linear combination of and , we obtain a channel-wise weighted feature map. Then we feed the channel-wise weighted feature map to the spatial attention model and obtain the spatial attention weights . After attaining two attention weights and , we can feed to modulate function to calculate the modulated feature map . All processes are summarized as follows:
where is a channel-wise multiplication for feature map channels and corresponding channel weights.
Spatial-Channel. The second type denoted as Spatial-Channel (S-C) is a model with spatial attention implemented first. For S-C type, given an initial feature map , we first utilize spatial attention to obtain the spatial attention weights . Based on , the linear function , and the channel-wise attention model , we can calculate the modulated feature following the recipe of C-S type:
where is an element-wise multiplication for regions of each feature map channel and its corresponding region attention weights.
We will validate the effectiveness of the proposed SCA-CNN framework for image captioning by answering the following questions: Q1 Is the channel-wise attention effective? Will it improve the spatial attention? Q2 Is the multi-layer attention effective? Q3 How does SCA-CNN perform compared to other state-of-the-art visual attention models?
4.1 Dataset and Metric
We conducted experiments on three well-known benchmarks: 1) Flickr8k : it contains 8,000 images. According to its official split, it selects 6,000 images for training, 1,000 images for validation, and 1,000 images for testing; 2) Flickr30k : it contains 31,000 images. Because of the lack of official split, for fair comparison with previous works, we reported results in a publicly available split used in previous work . In this split, 29,000 images are used for training, 1,000 images for validation, and 1,000 images for testing; and 3) MSCOCO : it contains 82,783 images in training set, 40,504 images in validation set and 40,775 images in test set. As the ground truth of MSCOCO test set is not available, the validation set is further splited into a validation subset for model selection and a test subset for local experiments. This split also follows . It utilizes the whole 82,783 training set images for training, and selects 5,000 images for validation and 5,000 images for test from official validation set . As for the sentences preprocessing, we followed the publicly available code 111https://github.com/karpathy/neuraltalk. We used BLEU (B@1,B@2, B@3, B@4) , METEOR (MT) , CIDEr(CD) , and ROUGE-L (RG) 
as evaluation metrics. For all the four metrics, in a nutshell, they measure the consistency between n-gram occurrences in generated sentences and ground-truth sentences, where this consistency is weighted by n-gram saliency and rarity. Meanwhile, all the four metrics can be calculated directly through the MSCOCO caption evaluation tool222https://github.com/tylin/coco-caption. And our source code is already publicly available 333https://github.com/zjuchenlong/sca-cnn.
In our captioning system, for image encoding part, we adopted two widely-used CNN architectures: VGG-19  and ResNet-152  as the basic CNNs for SCA-CNN. For the caption decoding part, we used an LSTM  to generate caption words. Word embedding dimension and LSTM hidden state dimension are respectively set to 100 and 1,000. The common space dimension for calculating attention weights is set to 512 for both two type attention. For Flickr8k, mini-batch size is set to 16, and for Flickr30k and MSCOCO, mini-batch size is set to 64. We use dropout and early stopping to avoid overfitting. Our whole framework is trained in an end-to-end way with Adadelta 
, which is a stochastic gradient descent method using an adaptive learning rate algorithm. The caption generation process would be halted until a special END token is predicted or a predefined max sentence length is reached. We followed the strategy of BeamSearch in the testing period, which selects the best caption from some candidates, and the beam size is set to 5. We noticed a trick that incorporates beam search with length normalization  which can help to improve performance in some degree. But for fair comparisons, all results reported are without length normalization.
4.3 Evaluations of Channel-wise Attention (Q1)
Comparing Methods. We first compared spatial attention with channel-wise attention. 1) S: It is a pure spatial attention model. After obtaining spatial attention weights based on the last conv-layer, we use element-wise multiplication to produce a spatial weighted feature. For VGG-19 and ResNet-152, the last conv-layer represents layer and , respectively. Instead of regarding the weighted feature map as the final visual representation, we feed the spatial weighted feature into their own following CNN layers. For VGG-19, there are two fully-connected layers follows layer and for ResNet-152, layer is followed by a mean pooling layer. 2) C: It is a pure channel-wise attention model. The whole strategy for the C type model is same as S type. The only difference is substituting the spatial attention with channel-wise attention as Eq. (4). 3) C-S: This is the first type model incorporating two attention mechanisms as Eq. (8). 4) S-C: Another incorporating model introduced in Eq. (9). 5) SAT: It is the “hard” attention model introduced in . The reason why we report the results of “hard” attention instead of the “soft” attention is that “hard” attention always has better performance on different datasets and metrics. SAT is also a pure spatial attention model like S. But there are two main differences. The first one is the strategy of modulating visual feature with attention weights. The second one is whether to feed the attending features into their following layers. All VGG results reported in Table 1 came from the original paper and ResNet results are our own implementation.
Results From Table 1, we have the following observations: 1) For VGG-19, performance of S is better than that of SAT; but for ResNet-152, the results are opposite. This is because the VGG-19 network has fully-connected layers, which can preserve spatial information. Instead, in ResNet-152, the last conv-layer is originally followed by an average pooling layer, which can destroy spatial information. 2) Comparing to the performance of S, the performance of C can be significant improved in ResNet-152 rather than VGG-19. It shows that the more channel numbers can help improve channel-wise attention performance in the sense that ResNet-152 has more channel numbers (i.e. 2048) than VGG-19 (i.e. 512). 3) In ResNet-152, both C-S and S-C can achieve better performance than S. This demonstrates that we can improve performance significantly by adding channel-wise attention as long as channel numbers are large. 4) In both of two networks, the performance of S-C and C-S is quite close. Generally, C-S is slightly better than S-C, so in the following experiments we use C-S to represent incorporating model.
4.4 Evaluations of Multi-layer Attention (Q2)
|Deep VS ||57.9||38.3||24.5||16.0||–||57.3||36.9||24.0||15.7||–||62.5||45.0||32.1||23.0||19.5|
|Google NIC ||63.0||41.0||27.0||–||–||66.3||42.3||27.7||18.3||–||66.6||46.1||32.9||24.6||–|
Comparing Methods We will investigate whether we can improve the spatial attention or channel-wise attention performance by adding more attentive layers. We conduct ablation experiments about different number of attentive layer in S and C-S models. In particular, we denote 1-layer, 2-layer, 3-layer as the number of layers equipped with attention, respectively. For VGG-19, 1-st layer, 2-nd layer, 3-rd layer represent conv-layer, respectively. As for ResNet-152, it represents conv-layer. Specifically, our strategy for training more attentive layers model is to utilize previous trained attentive layer weights as initialization, which can significantly reduce the training time and achieve better results than randomly initialized.
Results From Table 2 and 3, we have following observations: 1) In most experiments, adding more attentive layers can achieve better results among two models. The reason is that applying an attention mechanism in multi-layer can help gain visual attention on multiple level semantic abstractions. 2) Too many layers are also prone to resulting in severe overfitting. For example, Flickr8k’s performance is easier to degrade than MSCOCO when adding more attentive layers, as the size of train set of Flickr8k (i.e. 6,000) is much smaller than that of MSCOCO (i.e. 82,783).
4.5 Comparison with State-of-The-Arts (Q3)
Comparing Methods We compared the proposed SCA-CNN with state-of-the-art image captioning models. 1) Deep VS , m-RNN , and Google NIC  are all end-to-end multimodal networks, which combine CNNs for image encoding and RNN for sequence modeling. 2) Soft-Attention  and Hard-Attention  are both pure spatial attention model. The “soft” attention weighted sums up the visual features as the attending feature, while the “hard” one randomly samples the region feature as the attending feature. 3) emb-gLSTM  and ATT  are both semantic attention models. For emb-gLSTM, it utilizes correlation between image and its description as gloabl semantic information, and for ATT it utilizes visual concepts corresponded words as semantic information. The results reported in Table 4 are from the 2-layer C-S model for both VGG-19 and ResNet-152 network, since this type model always obtains the best performance in previous experiments. Besides the three benchmarks, we also evaluated our model on MSCOCO Image Challenge set c5 and c40 by uploading results to the official test sever. The results are reported in Table 5.
Results From Table 4 and Table 5, we can see that in most cases, SCA-CNN outperforms the other models. This is due to the fact that SCA-CNN exploits spatial, channel-wise, and multi-layer attentions, while most of other attention models only consider one attention type. The reasons why we cannot surpass ATT and Google NIC come from two sides: 1) Both ATT and Google NIC use ensemble models, while SCA-CNN is a single model; ensemble models can always obtain better results than single one. 2) More advanced CNN architectures are used; as Google NIC adopts Inception-v3  which has a better classification performance than ResNet which we adopted. In local experiments, on the MSCOCO dataset, ATT surpasses SCA-CNN only in BLEU4 and in METEOR, respectively. For the MSCOCO server results, Google NIC surpass SCA-CNN only in BLEU4 and in METEOR, respectively.
4.6 Visualization of Spatial and Channel-wise Attention
We provided some qualitative examples in Figure 3 for a better understanding of our model. For simplicity, we only visualized results at one word prediction step. For example in the first sample, when SCA-CNN model tries to predict word umbrella
, our channel-wise attention will assign more weights on feature map channels generated by filters according to the semantics like umbrella, stick, and round-like shape. The histogram in each layer indicates the probability distribution of all channels. The map above histogram is the spatial attention map and white indicates the spatial regions where the model roughly attends to. For each layer we selected two channels with highest channel-wise attention probability. To show the semantic information of the corresponding CNN filter, we used the same methods in. And the red boxes indicate their respective fields.
In this paper, we proposed a novel deep attention model dubbed SCA-CNN for image captioning. SCA-CNN takes full advantage of characteristics of CNN to yield attentive image features: spatial, channel-wise, and multi-layer, thus achieving state-of-the-art performance on popular benchmarks. The contribution of SCA-CNN is not only the more powerful attention model, but also a better understanding of where (i.e., spatial) and what (i.e., channel-wise) the attention looks like in a CNN that evolves during sentence generation. In future work, we intend to bring temporal attention in SCA-CNN, in order to attend features in different video frames for video captioning. We will also investigate how to increase the number of attentive layers without overfitting.
-  S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Lawrence Zitnick, and D. Parikh. Vqa: Visual question answering. In ICCV, 2015.
-  D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2014.
-  S. Banerjee and A. Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL, 2005.
-  K. Chen, J. Wang, L.-C. Chen, H. Gao, W. Xu, and R. Nevatia. Abc-cnn: An attention based convolutional neural network for visual question answering. In CVPR, 2016.
-  M. Corbetta and G. L. Shulman. Control of goal-directed and stimulus-driven attention in the brain. Nature reviews neuroscience, 2002.
-  J. Donahue, L. Anne Hendricks, S. Guadarrama, M. Rohrbach, S. Venugopalan, K. Saenko, and T. Darrell. Long-term recurrent convolutional networks for visual recognition and description. In CVPR, 2015.
-  H. Gao, J. Mao, J. Zhou, Z. Huang, L. Wang, and W. Xu. Are you talking to a machine? dataset and methods for multilingual image question. In NIPS, 2015.
-  K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. 2016.
-  S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 1997.
-  M. Hodosh, P. Young, and J. Hockenmaier. Framing image description as a ranking task: Data, models and evaluation metrics. JAIR, 2013.
-  X. Jia, E. Gavves, B. Fernando, and T. Tuytelaars. Guiding the long-short term memory model for image caption generation. In ICCV, 2015.
-  X. Jiang, F. Wu, X. Li, Z. Zhao, W. Lu, S. Tang, and Y. Zhuang. Deep compositional cross-modal learning to rank via local-global alignment. In ACM MM, pages 69–78, 2015.
-  A. Karpathy and L. Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015.
-  R. Krishna, Y. Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz, S. Chen, Y. Kalantidis, L.-J. Li, D. A. Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2016.
-  C.-Y. Lin. Rouge: A package for automatic evaluation of summaries. In ACL, 2004.
-  T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014.
M. Malinowski, M. Rohrbach, and M. Fritz.
Ask your neurons: A neural-based approach to answering questions about images.In ICCV, 2015.
J. Mao, W. Xu, Y. Yang, J. Wang, Z. Huang, and A. Yuille.
Deep captioning with multimodal recurrent neural networks (m-rnn).In ICLR, 2015.
-  V. Mnih, N. Heess, A. Graves, et al. Recurrent models of visual attention. In NIPS, 2014.
-  K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002.
-  M. Ren, R. Kiros, and R. Zemel. Exploring models and data for image question answering. In NIPS, 2015.
-  P. H. Seo, Z. Lin, S. Cohen, X. Shen, and B. Han. Hierarchical attention networks. arXiv preprint arXiv:1606.02393, 2016.
-  F. Shen, C. Shen, W. Liu, and H. Tao Shen. Supervised discrete hashing. In CVPR, pages 37–45, 2015.
-  F. Shen, C. Shen, Q. Shi, A. Van Den Hengel, and Z. Tang. Inductive hashing on manifolds. In CVPR, pages 1562–1569, 2013.
-  K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
-  M. F. Stollenga, J. Masci, F. Gomez, and J. Schmidhuber. Deep networks with internal selective attention through feedback connections. In NIPS, 2014.
-  C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818–2826, 2016.
-  R. Vedantam, C. Lawrence Zitnick, and D. Parikh. Cider: Consensus-based image description evaluation. In CVPR, 2015.
-  S. Venugopalan, M. Rohrbach, J. Donahue, R. Mooney, T. Darrell, and K. Saenko. Sequence to sequence-video to text. In ICCV, 2015.
-  S. Venugopalan, H. Xu, J. Donahue, M. Rohrbach, R. Mooney, and K. Saenko. Translating videos to natural language using deep recurrent neural networks. In NAACL-HLT, 2015.
-  O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption generator. In CVPR, 2015.
-  Y. Wei, W. Xia, M. Lin, J. Huang, B. Ni, J. Dong, Y. Zhao, and S. Yan. Hcp: A flexible cnn framework for multi-label image classification. TPAMI, 2016.
-  H. Xu and K. Saenko. Ask, attend and answer: Exploring question-guided spatial attention for visual question answering. In ECCV, 2016.
-  K. Xu, J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhutdinov, R. S. Zemel, and Y. Bengio. Show, attend and tell: Neural image caption generation with visual attention. In ICML, 2015.
-  Z. Yang, X. He, J. Gao, L. Deng, and A. Smola. Stacked attention networks for image question answering. In CVPR, 2016.
-  L. Yao, A. Torabi, K. Cho, N. Ballas, C. Pal, H. Larochelle, and A. Courville. Describing videos by exploiting temporal structure. In ICCV, 2015.
-  Q. You, H. Jin, Z. Wang, C. Fang, and J. Luo. Image captioning with semantic attention. In CVPR, 2016.
-  P. Young, A. Lai, M. Hodosh, and J. Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014.
-  M. D. Zeiler. Adadelta: an adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012.
-  M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014.
-  H. Zhang, Z. Kyaw, S.-F. Chang, and T.-S. Chua. Visual translation embedding network for visual relation detection. In CVPR, 2017.
-  Z. Zhao, H. Lu, C. Deng, X. He, and Y. Zhuang. Partial multi-modal sparse coding via adaptive similarity structure regularization. In ACM MM, pages 152–156, 2016.
-  Y. Zhu, O. Groth, M. Bernstein, and L. Fei-Fei. Visual7w: Grounded question answering in images. In CVPR, 2016.