Two-stream Spatiotemporal Feature for Video QA Task

TL;DR

This paper introduces a two-stream neural network with attention mechanisms for video question answering, leveraging spatiotemporal features to improve understanding of video content and answer accuracy.

Contribution

It proposes a novel multi-channel two-stream network with channel-wise attention and a context matching module for enhanced video question answering performance.

Findings

Improved results in textual-only setting on TVQA dataset.

Identified limitations and potential of visual features in the proposed model.

Demonstrated the effectiveness of attention mechanisms in joint spatiotemporal and textual modeling.

Abstract

Understanding the content of videos is one of the core techniques for developing various helpful applications in the real world, such as recognizing various human actions for surveillance systems or customer behavior analysis in an autonomous shop. However, understanding the content or story of the video still remains a challenging problem due to its sheer amount of data and temporal structure. In this paper, we propose a multi-channel neural network structure that adopts a two-stream network structure, which has been shown high performance in human action recognition field, and use it as a spatiotemporal video feature extractor for solving video question and answering task. We also adopt a squeeze-and-excitation structure to two-stream network structure for achieving a channel-wise attended spatiotemporal feature. For jointly modeling the spatiotemporal features from video and the textual features from the question, we design a context matching module with a level adjusting layer to remove the gap of information between visual and textual features by applying attention mechanism on joint modeling. Finally, we adopt a scoring mechanism and smoothed ranking loss objective function for selecting the correct answer from answer candidates. We evaluate our model with TVQA dataset, and our approach shows the improved result in textual only setting, but the result with visual feature shows the limitation and possibility of our approach.

Tables

Table 1. TABLE I: The list of commonly appearing notations.

Symbol	Description
$F_x$	The sequence of video frames. $x$ is a type of video frames (e.g. RGB or flow)
$V,V^S,V_x,V_x^S$	The spatiotemporal feature vectors from two-stream I3D. $V$ without the subscript $x$ is a set of spatiotemporal feature vectors. The superscript $S$ means the feature vectors with the time domain
$H_x$	The textual feature vectors. $x$ is a type of textual information (e.g. subtitles or query)
$G_x$	The context-aware query feature vector. $x$ is a type of query (e.g. $q$ is question and $a_i$ is $i$-th answer candidate)
$M_{x,i}$	A joint embedded feature vector. $x$ is a type of context embedded with the query (e.g. video or text), and $i$ means $i$-th answer candidate
$\odot$ and ;	An element-wise product and a concatenation arithmetic
$d$	A dimension of word embedding
$p_{x,i}$	An answer probability score for a fused feature vector between the context $x$ and $i$-th answer candidate.

Table 2. TABLE II: Accuracy of different methods on TVQA validation set. S, Q, and V indicate scripts, query, and video information, respectively.

Method	S+Q	V+Q	S+V+Q
Random	20.00	20.00	20.00
Lei et al. [3] w/ Image	65.15	43.78	66.44
Lei et al. w/ Region		44.40	67.17
Lei et al. w/ Concept		45.03	67.70
Our Image	66.01	36.00	64.84
Our RGB-I3D		35.72	58.25
Our Flow-I3D		35.54	58.63
Our two-stream I3D		35.96	58.55

Table 3. TABLE III: Accuracy of each question type when using baseline approaches and our approaches. Ours indicate two-stream I3D feature shown in Table  II . img is ImageNet features, reg is Faster-RCNN feature, cpt is the object detection label of Faster-RCNN, respectively.

	S+Q		V+Q				S+V+Q
	Lei et al. [3]	Ours	img	reg	cpt	Ours	img	reg	cpt	Ours
What(55.62%)	62.29	63.15	44.96	45.93	47.44	37.66	63.88	65.28	66.05	57.06
Who(11.52%)	68.33	69.19	35.75	34.85	34.68	23.53	67.76	67.20	67.99	64.28
Where(11.67%)	56.97	60.36	47.13	48.43	48.20	38.15	61.97	63.71	61.46	53.80
How(8.98%)	71.97	71.10	41.17	42.41	40.95	37.69	71.17	70.80	71.53	60.96
Why(10.38%)	78.65	79.14	45.36	45.36	45.48	37.73	78.33	77.13	78.77	63.23
Others(1.80%)	74.45	69.92	36.50	33.58	33.58	37.15	73.72	72.63	74.09	61.72

Table 4. TABLE IV: Comparison of accuracy of our methods w/ and w/o the SE block structure.

Method	V+Q	S+V+Q
RGB-I3D w/o SE	35.44	53.16
Flow-I3D w/o SE	35.78	54.16
Two-stream I3D w/o SE	35.85	53.91
RGB-I3D w/ SE	35.72	58.25
Flow-I3D w/ SE	35.54	58.63
Two-stream I3D w/ SE	35.96	58.55

Equations

$$s=F_{ex}(z,W)=\sigma (g(z,W))=\sigma (W_2\delta (W_{1}z)),$$

$$F_{scale}:{\mathbb{R}}^{H\times W\times L\times C}\times {\mathbb{R}}^{1\times 1\times 1\times C}\to {\mathbb{R}}^{H\times W\times L\times C},$$

$$\tilde{X}=F_{scale}(U,s)=U\odot s,$$

$$G$$

$$S$$

$$M_{video,i}=[V^{{}^{\prime }S};G_q;G_{a_i};V^{{}^{\prime }S}\odot G_q;V^{{}^{\prime }S}\odot G_{a_i}]$$

$$M_{video,i}\in {\mathbb{R}}^{n_{video}\times 2000},$$

$$l_{LSEP}=\log \left(1+\sum _{v\notin Y_i}\sum _{u\in Y_i}\exp (p_{x,v}-p_{x,u})\right),$$

Full text

I Introduction

Understanding the content of videos is one of the main tasks in the computer vision field. This is mainly because this understanding task of videos can serve as a core technology for developing various useful applications in the real world. For example, recognizing various human actions can be useful for surveillance systems (e.g., detecting criminal actions) or customer behavior analysis in an autonomous shop. Furthermore, understanding the story of a video also can be applied to assisting visually impaired people or the interactive educational system for children.

Unfortunately, understanding the content or story of the video itself still remains a challenging problem because of its very nature. Since the video consists of a sequence of images, it has a temporal structure that does not exist in a single image. As a result, techniques targeting the video understanding should not only deal with the challenges in the spatial domain such as background clutter or object occlusion but also deal with the challenges in the temporal domain such as moving viewpoint or reasoning the order of different events.

Thanks to the progress of deep learning technique, there have been several deep learning based approaches for video related tasks such as video question and answer (QA) [1, 2, 3, 4, 5, 6, 7, 8, 9], or human action recognition [10, 11, 12]. The approaches for solving the video QA task generally adopts LSTM [13], GRU [14] or C3D [15], as a spatiotemporal video feature extractor that gets a sequence of RGB images as input. On the other hand, approaches for solving human action recognition adopt a two-stream ConvNet structure [10], which gets RGB and optical-flow as input to its spatiotemporal video feature extractor. Interestingly, it has been known that for human action recognition [10, 12, 11], the two-stream ConvNet spatiotemporal extractor shows higher accuracy than those of the single-stream ConvNet or recurrent neural network (RNN) spatiotemporal extractor.

Inspired by the advances in this field of human action recognition, we propose a multi-channel neural network structure with two-stream I3D [11] spatiotemporal feature extractor for solving the video QA task. We also adopt the Squeeze-and-Excitation(SE) structure [16] to the two-stream I3D to apply a channel-wise attention mechanism and make the network concentrate on important objects and actions in the video. To jointly modeling the spatiotemporal features from video and textual features from the question, we also design a level adjusting layer to remove the gap of information levels between two of them. We adopt a scoring mechanism for selecting the correct answer from answer candidates and use a smoothed ranking loss LSEP [17] as an objective function.

We evaluate our model with the TVQA dataset [3], which provides the sequence of video frames with subtitle as context. The dataset also provides a question and its answer candidates as the query. Our approach shows the improved result in a textual only model, but the result with the visual model shows the limitation and possibility simultaneously.

II Related Work

II-A Video QA task

Video question and answering (QA) task is a challenging computer vision task where a computer needs to answer questions given with input videos. Though challenging, the task is worth studying since it is an effective way to evaluate how well a model understands the content of videos; we can form any kinds of questions to test our QA models, from naive ones (e.g., what, where, etc.) to more profound ones (e.g., how, why, etc.).

Most video QA methods have included subtitles or scripts ,as well as the visual cue in that actors’ lines in the text, are crucial to grasp essential information on videos. Therefore, to solve the video QA task, the system needs to extract proper features from both visual inputs (i.e., RGB frames and optical flow) and textual inputs (i.e., subtitle, query, and answer candidates), and adequately correlate those features to infer correct answers. When compared to image QA task, video QA task has more challenges in that it needs to additionally deal with the temporal domain of visual information and connect each feature from different modalities temporally. Even setting aside the multi-modality, extracting good visual features rich in temporal information itself is difficult and has been actively studied in the field of video recognition.

Researchers have been approached video QA task from various perspectives. Na et al. [1] and Kim et al. [2] propose a deep model based on memory network architectures for embedding the story of videos and reasoning the correct answer. Zhu et al. [6] adopt a GRU encoder-decoder for modeling the temporal structure of a video and apply a scoring mechanism for choosing the correct answer. Various techniques [7, 8, 9, 4] adopt a spatiotemporal attention mechanism to select important features from the appearance and motion information to solve the questions. Also, 3D ConvNet [7, 8, 9] is commonly used for extracting temporal features from RGB video frames.

Many previous approaches use the ImageNet [18] pre-trained network for extracting spatial features and use LSTM [13], GRU [14], or C3D [15] for extracting temporal features from the sequence of videos. However, it has been shown for action recognition tasks that two-stream method [10, 12, 19] that utilizes optical flows for temporal cues has been more successful in terms of video understanding than other methods. Therefore, departing from previous video QA work, we adopt a two-stream network structure for extracting useful spatiotemporal features from the sequence of video frames.

II-B Two-stream network structure

Thanks to its strength in processing the spatiotemporal domain, the two-stream network structure has been widely used in the action classification field. Simonyan et al. [10], Wang et al. [12], and Fan et al. [19] use a two-stream ConvNet that gets two kinds of inputs: one is a single frame of a video for the spatial stream ConvNet, and the other is a multi-frame optical flow of the video for the temporal stream ConvNet.

The two-stream network structure suggested by Carreira et al. [11] also gets two kinds of inputs but takes both the sequence of RGB and optical flow, respectively. Simonyan et al. [10], Wang et al. [12], and Carreira et al. [11] show higher accuracy on action recognition tasks over single-stream architectures or recurrent neural networks in dealing with the temporal domain of videos. Fan et al. [19] show its capability of processing spatiotemporal domain features with identifying a camera wearer from a third-person view camera scene.

In this work, we propose to use the two-stream ConvNet for video QA task, focusing on its ability to process spatiotemporal domain features.

II-C Attention mechanism

Attention mechanism has been widely used for various applications including image search [20, 21, 22]. Since the queries of the video QA task generally ask about a specific object or event at a specific timing in a story, solving the video QA task needs to focus on the important information that is closely related to the queries from the story.

Seo et al. [23] present an attention flow layer that makes both context-aware query and query-aware context vectors by computing a similarity matrix and using it as an attention mask. Lei et al. [3] adopt the attention flow layer as a context matching module and feed the context-aware vectors into bidirectional LSTM for jointly modeling the context and query.

Departing from the vector-wise attention methods, Hu et al. [16] present a channel-wise attention method, Squeeze-and-Excitation (SE) structure, that can be applied to any given transformation $F_{tr}:X\mapsto U$, where $X\in\mathbb{R}^{H^{{}^{\prime}}\times W^{{}^{\prime}}\times C^{{}^{\prime}}},U\in\mathbb{R}^{H\times W\times C}$. SE block can be easily attached to the existing ConvNet models such as ResNet [24] or GoogLeNet [25], and the ConvNet with SE block shows a better result than ConvNet without SE block at the image classification task.

In this work, we apply the SE structure to our two-stream network for extracting features with the channel-wise attention of the spatiotemporal domain. We also utilize the context matching module for matching the spatiotemporal features from the video frames and the queries.

III Our Approach

In this study, we propose a multi-channel neural network structure with two-stream spatiotemporal video feature extractor for solving multimodal video QA task. Our model includes a set of video and subtitle as context and a set of question and answer candidates as the query. Specifically, we focus on a method for processing the video context with our novel approach, which extracts the spatiotemporal features of videos with the channel-wise attended two-stream network structure.

Our approach, which is shown in Figure 1, is largely divided into two parts of the neural network stream: visual stream and textual stream. Each stream starts with corresponding feature extractors: one for video consists of two-stream I3D [11] and the other one for text has GLoVe [26] followed by bi-directional LSTM. Both streams go through several additional layers including context matching, fusion and scoring layer; each stream has subtly different components and the details will be discussed in the next following sub-sections. Each input (i.e. video or text) processed via several stages produces predictions for the answer. We aggregate the predictions at the final step to put all information from different modalities together and get the definitive answer.

In the next following sub-sections, we discuss the details of our approach. For clarity, we summarize common notations used throughout the paper in Table I.

III-A Two-stream I3D with SE structure

The visual stream of our method starts with the feature extraction stage. In this work, we adopt two-stream I3D [11], which shows its capacity for processing video frames in action classification task. However, unlike simple classification task, visual features for QA task are highly required to focus on salient objects and disregard others since it has to be correlated (i.e. context matching) to textual features which are relatively more focused on necessary context by its nature. Therefore, we adapt Squeeze-and-Excitation (SE) [16] structure for temporal inputs and add it to several layers of I3D extractor to generate more refined and attended visual features.

Our visual feature extractor (see Fig. 2) is based on I3D pretrained on ImageNet [18] and Kinetics [27] dataset. It produces a tuple of visual features $V^{S}=\{V_{spt}^{S},V_{tpr}^{S}\}$ which includes spatial feature $V_{spt}^{S}$ from RGB frames $F_{RGB}=\{a_{0},a_{1},...,a_{n}\}$ and temporal feature $V_{tpr}^{S}$ from flow frames $F_{flow}=\{b_{0},b_{1},...,b_{n-1}\}$, where $n$ is the number of frames in a sequence, $a_{i}\in\mathbb{R}^{224\times 224\times 3}$ and $b_{i}\in\mathbb{R}^{224\times 224\times 2}$. Here, we need to have temporal sequence preserved in extracted features since in context matching stage, the features are temporally matched and attended with query features. Therefore, different from I3D [11] producing two 400-dimensional vectors from RGB and flow (i.e. $\{V_{spt},V_{tpr}\}$, where $V_{spt},V_{tpr}\in\mathbb{R}^{400}$ ), we remove temporal pooling layers to preserve the temporal sequence so that we get $V_{spt}^{S}\in\mathbb{R}^{n\times 400}$ and $V_{tpr}^{S}\in\mathbb{R}^{(n-1)\times 400}$; the preserved temporal sequence is utilized in the context matching phase.

The video frames included in the video QA task commonly have unnecessary information (e.g. background clutters or unrelated objects) which degrades the performance of a video QA model. The textual queries and subtitles, however, have less clutters because textual information is usually well focused on related objects or actions. This kind of gap between two modalities raise a difficulty in context matching stage where visual and textual information is merged.

To make our feature extractor concentrate more on the crucial objects in the video frames, we utilize the Squeeze-and-Excitation (SE) structure [16] and integrate it within the two-stream I3D, where a feature vector $X$ is transformed into another feature vector $U$ by the inception module [25, 11]. Fig. 2 shows how Sqeeuze-and-Excitation structure is merged into the visual feature extractor.

In the SE block structure of two-stream I3D, the Squeeze operation embeds global spatiotemporal information of output feature vectors of the Inception module into a channel descriptor, $z\in\mathbb{R}^{1\times 1\times 1\times C}$, with global average pooling (Fig. 3). The Excitation operation generates the attended channel descriptor $s$ by two fully-connected layers with activation functions:

[TABLE]

where $z$ is the squeezed feature vectors, $\sigma$ is the sigmoid function, $\delta$ is the ReLU [28] function, and $W_{1}\in\mathbb{R}^{{C\over r}\times C}$ and $W_{2}\in\mathbb{R}^{C\times{C\over r}}$ are the weight parameters of the fully-connected dimensionality-reduction layer and dimensionality-increasing layer, respectively. The FC layers include a parameter $r$ for the ratio of dimensionality reduction.

By introducing the scale function:

[TABLE]

which operates on the channel descriptor $s$ and the output feature vector of the Inception block $U$, we get the channel-wise attended feature vector $\tilde{X}$:

[TABLE]

where $\odot$ is the element-wise product with shape broadcasting. The SE block structure can be added to each inception module without changing the core structure of I3D to produce better features.

III-B Multimodal joint embedding and context matching

To better express the needed visual context conditioned by queries, extracted spatiotemporal features should be correlated and jointly embedded with them. For the multimodal joint embedding, we adopt an attention flow layer proposed by Seo et al. [23] and Lei et al. [3]. By the method, the spatiotemporal feature vectors and the textual feature vectors of the query are jointly embedded to form a set of context-aware feature vectors.

We use GLoVe [26] word embedding to vectorize each word in the query. Query sentences composed of multiple GLoVe vectors are processed by bidirectional LSTM [29, 13] which constructs our textual features $H_{query}\in\mathbb{R}^{n_{query}\times 2d}$, from the sequence of words in the query, where $n_{query}$ is the number of words in query consisting of the question and the answer candidates.

Although visual features attended by SE blocks, visual I3D feature vectors $V^{S}$ and textual GloVe vectors $H_{query}$ have a different level of information, and there is also innate domain gap between them. This gap of information level can disturb the joint embedding of multimodal features and deliver the obscure information to the prediction layer.

For these reasons, we put an information level adjusting layer to remove the gap of information level between the two types of feature vectors. The adjusting layer consists of a learnable fully-connected layer to each feature vectors that have $W\in\mathbb{R}^{400\times 400}$ as weight and Leaky ReLU [30] as an activation function.

Our information-level adjusting layer produces a set of level-adjusted spatiotemporal feature vectors $V_{spt}^{{}^{\prime}S}\in\mathbb{R}^{n_{RGB}\times 400},V_{tpr}^{{}^{\prime}S}\in\mathbb{R}^{n_{flow}\times 400}$ and the calibrated textual query feature vectors $H_{query}^{{}^{\prime}}\in\mathbb{R}^{n_{query}\times 400}$.

The calibrated spatiotemporal feature vectors of video frames and the textual query feature vectors are jointly modeled to produce a context-aware query feature $G\in\mathbb{R}^{n_{video}\times 400}$ in a context matching layer [23, 3]:

[TABLE]

where $S\in\mathbb{R}^{n_{video}\times{n_{query}}}$ is a similarity matrix. We calculate the similarity matrix $S$ with the matrix multiplication operation to link each sequence of video feature vectors and query feature vectors, and the softmax function to emphasize the important information in the similarity matrix.

The calibrated spatiotemporal feature vectors $V^{{}^{\prime}S}$ and the context-aware query feature vector $G$ are fused together to form a multimodal joint embedded feature vector $M_{video,i}=\{M_{spt},M_{tpr}\}$:

[TABLE]

where $n_{video}$ is the number of RGB or optical flow frames, $q$ is a question and $a_{i}$ is the $i$-th answer candidate. The dimension of $M_{video,i}$ is $n_{video}$ times 2,000, since we concatenate five vectors, each of which is in 400 dimensions.

To predict the correct answer from $M_{video,i}$, we convert the fused feature vectors $M_{video,i}$ to scalars of probability score ${p_{video,i}}=\{p_{spt},~{}p_{tpr}\}$ with an FC layer and a temporal max-pooling layer, which can choose the most important information from the spatiotemporal-fused feature.

III-C Processing of textual context and answer prediction

We also perform the joint modeling of the query and the context encoded in another textual context (e.g. subtitles), which is already included in the video QA task. The textual context is processed by bidirectional LSTM [29, 13] and fused with the query feature vectors to form the joint embedded feature $M_{text,i}$. The fused feature $M_{text,i}$ is encoded with bidirectional LSTM again to extract the temporal information and max-pooled in the temporal domain to get the answer probability score ${p_{text,i}}$.

Finally, we normalize the answer probability score $p_{video,i}$ and $p_{text,i}$ with the softmax function and sum up to get the final answer probability score. To make the correct answer candidate’s score is higher than any other wrong answer candidates, we adopt the log-sum-exp pairwise (LSEP) function [17], which is a smooth approximation of the marginal hinge ranking loss, as an object function:

[TABLE]

where $Y_{i}$ is the correct answer. By this smoothed ranking loss function, we can pose a margin between the wrong answer candidates and the correct answer candidates in the feature space.

IV Experiment

In this section, we evaluate the effectiveness of our two-stream spatiotemporal video feature extractor in a multi-channel neural network structure. We tested our model on the TVQA dataset [3] and compared our method against several baseline approaches. Finally, we analyzed the prediction accuracy of our model on different types of questions in the validation set.

IV-A Dataset

The TVQA dataset [3] includes 152,545 QA pairs from 21,793 TV show clips. The QA pairs are split into the ratio of 8:1:1 for training, validation, and test sets. The TVQA dataset provides the sequence of video frames extracted at 3 FPS, the corresponding subtitles with the video clips, and the query consisting of a question and four answer candidates. Among the four answer candidates, there is only one correct answer.

Since the TVQA dataset provides only the sequence of RGB video frames for the visual context, we computed the optical flow frames with TV-L1 algorithm [31] for our two-stream spatiotemporal feature extractor. The dataset also provides the timestamps for each query, so we trained and tested our model with the timestamps to localize the video and subtitle data.

IV-B Implementation details

We train the textual context processing channel and the visual context processing channel separately for reducing the variances of neural networks and achieving the effectiveness of the model ensemble [32].

When training the textual channel, we use the Adam optimizer [33], where an initial learning rate is 0.0003, a momentum parameter $\beta_{1}$ is 0.9, a momentum parameter $\beta_{1}$ is 0.999, numerical stability parameter $\epsilon$ is $1e^{-8}$, and an exponential decay rate for 0.9 at every five epochs. The model is trained for 100 epoch with the early stopping method [34], where the patience value is three, for preventing the overfitting problem. We train our textual channel on a machine, which has Intel Xeon CPU E5-2650 v4 @ 2.20GHz, 64GB of RAM, and four Nvidia GTX 1080Ti GPU. The mini-batch sizes of each GPU are set to 32, and we take five days for training.

When training the visual channel, we import the pre-trained two-stream I3D [11], which was trained with the Kinetics dataset [27] that includes 25fps videos, and then fine-tune it with the TVQA dataset. We use two kinds of optimizers, which are Adam [33] and SGD [35], because the two-stream I3D feature extractor empirically requires a higher learning rate [11] than other layers (e.g. context matching layer or calibration layer) in the visual channel. We set the maximum number of frames to 69 per question due to the limitation of VRAM in our GPU. We use SGD optimizer [35] for training the two-stream I3D spatiotemporal feature extractor, where an initial learning rate is 0.02, a momentum parameter for 0.9, and an exponential decay rate for 0.9 at every five epochs. Each stream of two-stream I3D is trained separately, and their predictions are combined at the inference time.

For training the calibration layer and scoring layer in the visual channel, we use the Adam optimizer, where an initial learning rate is 0.0003, a momentum parameter $\beta_{1}$ is 0.9, a momentum parameter $\beta_{1}$ is 0.999, numerical stability parameter $\epsilon$ is 0.1, and an exponential decay rate for 0.9 at every five epochs. The model is trained for 40 epochs with the early stopping method. A machine which we train our visual channel with has Intel Xeon CPU E5-2630 v4 @ 2.20GHz, 1TB of RAM, and eight Nvidia RTX 2080Ti GPU. The mini-bach sizes of each GPU are set to 4 and it takes three weeks for training. Our model is implemented with Tensorflow [36].

IV-C Experiment results on TVQA dataset

Table II shows the results from baseline methods to our model on the experiment of TVQA dataset. All experiments are tested with the timestamp option in the dataset, which is used for localizing the video and subtitle related to the query.

We test three baseline results adopted from the work of Lei et al. [3], which are tested with three kinds of video features: image indicates the ImageNet [18] pre-trained ResNet101 [24] features, which are extracted from convolutional layer5 after pooling and has 2048 dimension, region uses the pre-trained Faster-RCNN [37, 38] features, and concept uses the detected object labels of the pre-trained Faster-RCNN.

Our Image, RGB-I3D, Flow-I3D, and Two-stream I3D shown in Table II are variations of our model: Image uses the pre-trained ResNet101 features, whose dimension is reduced to 400 for fitting to our model, RGB-I3D uses the spatial video features from the sequence of video RGB frames, Flow-I3D uses the temporal video features from the sequence of optical-flow frames, and Two-stream I3D uses both spatiotemporal video feature. The random selection model gets 20% accuracy because the TVQA dataset has one correct answer among five answer candidates.

When we evaluate the models with subtitle and query information, our model gets better results than the tested baseline methods. We assume that the LSEP loss function, which we adopt instead of cross-entropy loss, helps our model to learn better features and solve more difficult questions than the baseline model.

We expected that the features from the two-stream network show higher accuracy than using the ImageNet feature, but the ImageNet feature from ResNet101, Our Image, gets the highest accuracy when considering both video and query among the tested four different types of video features evaluated with our model. Furthermore, the ImageNet feature in our model shows a lower result than the baseline methods with the ImageNet feature. This is caused by the reduction of dimension on the ImageNet feature to fit our model to the context matching, degrading the amount of information in the feature.

Under the test setting of S+V+Q, we find that all kinds of video features show lower accuracy than the baseline, and they even degrades the performance of S+Q setting. Especially, the video features from I3D much degrades the performance of S+Q setting than the ImageNet feature (Our Image). To find out why our video features get an inferior result over the baseline and degrade the accuracy of the text-only setting, we analyze the baseline and our approaches’ accuracy of each question type.

Table III shows the accuracy of different methods under each question type. We split the questions into six types of what, who, where, how, why, and others. As shown in S+Q column, our text only model improves accuracy in every question types except for ’why’ and ’others’ which account for 10% of the total questions. Specifically, our model improves the most in ’where’ questions, where the original model shows the lowest accuracy.

On the contrary to the result of S+Q, our model shows lower accuracy than baseline methods in every question types except for ’others’ in V+Q. Our hypothesis of this result is that our spatiotemporal feature extractor with the two-stream I3D has difficulty in extracting the feature from the provided video data, because 3fps, which is a frame rate of provided video data, is too low for extracting sufficiently dense optical-flow for our two-stream I3D, which is originally trained on 25fps videos. Nevertheless, our V+Q model shows the highest accuracy on ’where’ question type, which our S+Q model shows the lowest accuracy. This result shows the possibility that our visual model can complement the textual model.

IV-D The effectiveness of the SE structure

To see the effectiveness of our SE structure, we perform ablation study w/ and w/o the SE block structure in Table IV. We find that the models with the SE structure show higher accuracy than the models without the SE structure in S+V+Q. Based on this result, we can see that the SE structure in our method helps to extract the complementing spatiotemporal features to the textual feature.

V Conclusion

In this paper, we have proposed a multi-channel neural network structure for solving the video QA task. We use two-stream I3D for extracting temporal features from the sequence of video frames, and apply the SE structure between the Inception blocks in the two-stream I3D. The SE structure gives the effect of channel-wise attention to the network, so it makes the network to focus on the important objects in the video. For multimodal joint embedding of the video and text features, we modify two-stream I3D to produce the features with the sequence in it, and we also design the information-level adjusting layer to reduce the gap of information levels between the two types of features.

To evaluate our approaches, we conducted the experiment with TVQA dataset and ablation study with the SE structure. Our approach showed the improved result in the textual model, but the result with the visual model showed its limitation with possible future research directions. The ablation study with SE structure showed the effectiveness of the SE structure that makes the visual model to complement the textual model.

V-A Limitations and future work

Since the visual channel in our model spent too much time for training, we needed more than a week to check the experiment result. This heavy requirement of computational power led to the insufficient amount of attempts for finding a proper architecture design and hyper-parameters.

In future work, we will train and evaluate our model in a more efficient way and search for the cause of the malfunction in the visual channel, such as modifying the structure of I3D to work with low frame videos or checking the context module whether it works properly to reduce the gap of information level.