Predicting Group Cohesiveness in Images

TL;DR

This paper introduces a neural network-based approach to estimate group cohesiveness from images, combining user surveys and novel databases, achieving near human-level accuracy and revealing a correlation with group emotion.

Contribution

It proposes a multi-task CNN and capsule network for predicting group cohesion, introduces the GAF-Cohesion database, and demonstrates the link between group emotion and cohesion.

Findings

Model achieves near human-level performance

Group cohesion correlates with group emotion

Introduces the GAF-Cohesion database

Abstract

The cohesiveness of a group is an essential indicator of the emotional state, structure and success of a group of people. We study the factors that influence the perception of group-level cohesion and propose methods for estimating the human-perceived cohesion on the group cohesiveness scale. In order to identify the visual cues (attributes) for cohesion, we conducted a user survey. Image analysis is performed at a group-level via a multi-task convolutional neural network. For analyzing the contribution of facial expressions of the group members for predicting the Group Cohesion Score (GCS), a capsule network is explored. We add GCS to the Group Affect database and propose the `GAF-Cohesion database'. The proposed model performs well on the database and is able to achieve near human-level performance in predicting a group's cohesion score. It is interesting to note that group cohesion as an attribute, when jointly trained for group-level emotion prediction, helps in increasing the performance for the later task. This suggests that group-level emotion and cohesion are correlated.

Tables

Table 1. TABLE I: GCS and emotion recognition comparison.

GAF 3.0	Ours	VGG 16	AlexNet
MSE, GCS	0.8181	0.8967	1.0375
Accuracy(%) Group Emotion	85.58	40.26	72.21

Table 2. TABLE II: Image-level network architecture. Here, b and BN refer to the batch size and batch normalization respectively.

Layers	Input	Output	Layer Details
Inception V3	b,224,224,3	b,2048	similar to [30]
Dense	b,2048	b,4096	4096
Activation	b,4096	b,4096	Relu/Swish
Dense	b,4096	b,4096	4096
Activation	b,4096	b,4096	Relu/Swish
Dense	b,4096	b,4096	4096
Activation	b,4096	b,4096	Relu/Swish
Cohesion (Sigmoid)	b,4096	b,1	1

Table 3. TABLE III: Comparison of the performance of CapsNet with other networks on RAF-DB.

RAF-DB	Ours	Alexnet	mSVM [34]	DLPCNN [34]
Accuracy(%)	77.48	76.27	65.12	74.20

Table 4. TABLE IV: Face-level Network Architecture. Here, b and BN refer to the batch size and batch normalization respectively.

Layers	Input	Output	Layer Details
Dense	b,3,7	b,3,16	16
BN and Activation	b,3,16	b,3,16	Relu/Swish
Dense	b,3,16	b,3,32	32
BN and Activation	b,3,32	b,3,32	Relu/Swish
Max Pooling	b,3,32	b,1,32	3(1-D)
Flatten	b,1,32	b,32	-
Cohesion (Sigmoid)	b,32	b,1	1

Table 5. TABLE V: The results of image-level group emotion (classification accuracy) and cohesion (MSE) analysis. Notice that the performance of group emotion increases, when it is jointly trained with AGC. However, the same is not true for AGC .

Network	Accuracy (%)	MSE
Inception V3 (emotion and cohesion prediction)	85.58	0.8181
Inception V3 (emotion prediction)	65.41	NA
Inception V3 (cohesion prediction)	NA	0.8537

Table 6. TABLE VI: Comparisons of GCS prediction using the image-level and face-level networks. Due to copyright we will release a subset of the data for the EmotiW 2019 challenge. The baselines are mentioned in the last column of this table.

Network Details	Image-Level	Face-level	EmotiW Baseline
GCS (MSE on Val. set)	0.85	1.11	0.84
GCS (MSE on Test set)	0.53	0.91	0.50

Table 7. TABLE VII: 5 fold cross validation results of the GAF cohesion database. lr = learning rate

Cross validation	MSE (lr=0.001)	MSE (lr=0.01)
1st	0.63958	0.65662
2nd	1.10628	1.06666
3rd	0.70162	0.67964
4th	0.60604	0.76320
5th	0.93969	0.89159
Average	0.79864	0.81155

Table 8. TABLE VIII: Group emotion performance comparison on GAF.

GAF 2.0	Ours	[36]	[26]	[25]	[13]
Accuracy(%)	85.67	83.90	80.05	77.92	52.97

Full text

I Introduction

The concept of ‘teamwork’ is defined as the collaborative effort of a group of people to accomplish a common goal in the most well-organized way [1]. One of the most important requirements for effective teamwork is cohesion. The main motivation of our work is to understand the human perception of Group Cohesiveness Score (GCS) [2] from images and map the attributes to an Automatic Group Cohesion (AGC) pipeline. Group cohesiveness is defined as the measure of bonding between group members. Higher cohesiveness implies stronger group-level bonding. According to psychological studies, group cohesion depends on several factors such as members’ similarity [3], group size [4], group success [5] and external competition and threats [6, 7]. The reason behind a strong group bonding can be positive (e.g. group success) or negative (e.g. threats). One of the key factors behind any group-level success is high group cohesiveness [8] as it affects group-level performance. Beal et al. [8] argue that group cohesion plays the most important role in group performance. Similarly, group members’ satisfaction [9] also plays an important role in deciding the cohesiveness of a group.

Hackman et al. [9] state that members belonging to a cohesive group have more satisfaction as compared to a non-cohesive group. Myers [10] indicates that people belonging to a cohesive group are less prone to anxiety and tension. Lott et al. [11] found that group cohesion helps improve individual members’ learning processes. Inspired by the aforementioned studies, in this work we are interested in investigating the following research questions:

• •

How useful are holistic (image-level) and facial information for predicting cohesion in a group?

• •

What are the factors that affect the perception of the cohesiveness in a group?

• •

What is the usefulness of cohesiveness as an attribute for tasks such as group emotion prediction?

In this work, we investigate AGC analysis from an early prediction perspective. This can also be viewed as a first impression of a group’s cohesion, similar to the early personality assessment [12] problem in affective computing. The main contributions of this paper are as follows:

*To the best of our knowledge, this is the first study proposing AGC prediction in images; * 2. 2.

We compare two cohesion models, representing scene (holistic) and face-level information respectively, and show that the former contributes more to the perception of cohesion; 3. 3.

*We label and extend the Group Affect Database [13] with group cohesion labels and propose the GAF Cohesion database (sample images from the database are shown in Fig. 1); * 4. 4.

From our experimental results, we observed that the perceived group emotion is related to group cohesiveness (Section VI).

The rest of the paper is structured as follows: Section II describes the prior works on Group cohesion. Section III explains the challenges involved in predicting the GCS task and the procedure of our survey. Section IV discusses the data and labeling process. The details of the proposed methods are described in Section V. Experiments are discussed in Section VI. Section VII describes regarding the visual attributes that our network learn. Conclusion, limitations and future research directions are discussed in Section VIII.

II Prior Work

II-A Group-level Cohesion (Psychological Aspects)

According to Barsade et al. [14], several factors impact the perception of a group’s cohesion and emotion. The authors [14] argued that social norms and constraints (i.e. interpersonal bonding and individual emotional responses) are important cues for group emotion and cohesion. Gallagher et al. [15] modelled the group as a min span tree based on facial locations and inferred the gender and age of group members using the group-level contextual information. Tajfel et al. [3] stated that one of the main factors which affect a group’s cohesiveness is its group members’ similarity. Here, similarity can be measured in terms of their occupation, ethnicity, age and relationship etc. This may also imply that due to these factors group members may have a similar point of view about certain issues, which may cause strong bonding between them. Another interesting study by Carron et al. [4] suggested that a small group implies strong cohesion. The reason behind this is that as the number of group members increases, their opinions may vary. This may lead to weaker cohesiveness as compared to small groups. Zaccaro et al. [5] argued that group-level success (towards a task) is another factor, which influences cohesiveness, along with the group’s size and its members’ similarity. Apart from the positive factors, some negative factors may also influence a group’s cohesiveness. Several studies [6, 7] revealed that threats to a group and competition with another group may also increase a group’s cohesiveness.

In a seminal work, Hung et al. [16] studied group cohesion in a constrained environment using audiovisual-based group meeting data. Several audio and video features were extracted to test their importance on group cohesion. For audio analysis, pauses between individual turns, pauses between floor exchanges, turn lengths, overlapping speech, prosodic cues etc. are taken into consideration. Similarly, video features include pauses between individual turns, pauses between floor exchanges, motion turn lengths, overlapping visual activity, visual energy cues, ranking participants’ features and group distribution features etc. Further, an SVM based classifier is used for predicting overall cohesion score. To the best of our knowledge, this is the first work that investigates the automatic cohesion of a group of people in videos.

II-B Study of ‘Group of People’

In recent years, computer vision researchers have studied automatic analysis of ‘group of people’ for different tasks. In an interesting work, Chang et al. [17] predicted group-level activity via hierarchical agglomerative as well as the divisive clustering algorithm. In order to track the group-level activity, multiple cameras are placed in different environments (e.g. in an abandoned prison yard) which first detect group related information such as group formation, dispersion and distinct groups. Further, it investigates motion patterns (Loitering, Fast Moving, Approaching, Following) and behaviour (e.g. Flanking, Agitation, Aggression). In another work, Wang et al. [18] proposed a method to infer the relationship between group members via geometric structure and appearance based features of the group. The AMIGOS database [19] has been recently proposed to study different aspects of affect in a group-level setting.

II-C Group Emotion

One of the first group emotion analyses was proposed by Dhall et al. [20]. They proposed the Group Expression Model (GEM) to predict happiness intensity of a group of people in images. Several other studies [21, 22, 23, 24, 25, 26, 27] mainly extracted scene, face and pose features to predict group emotion. In another recent paper, Singh et al. [28] studied the effect of a group on a person’s smile. They evaluated the usefulness of visual features in predicting the task.

III Challenges and Survey Result

This section describes the challenges involved in designing an AGC prediction network. To design an automated system for AGC prediction, we wish to know the factors which affect the perception of cohesion of a group. In the existing literature, the perception of members’ similarity [3] is claimed to be a vital visual cue; however, the first perception after viewing an image differs considerably from person to person. To understand the important visual cues, we have conducted a survey. The survey is conducted via Google form with 102 participants. There are 59 male and 43 female participants belonging to age group 22-54 years. The participants are from different backgrounds like student, businessman, corporate employee etc. The form consists of 24 images (as shown in Fig. 3) of groups of people in different contexts and having different GCS values (6 images for each GCS value). Based on Treadwell et al. [2], we use four levels of cohesion. Before filling in the form, the participants are familiarized with the concept of group cohesion labels [2] with images. The participants have to select one of the four cohesion levels for each image and they have to provide reasons behind their choice. Thus, we are provided some keywords related to the AGC score and corresponding image. After analyzing the responses, we get the statistics as shown in Fig. 5. From the word clouds of Fig. 5, we can see that ‘team’, ‘bonding’ and ‘together’ are the most frequent keyword responses which indicate that we are dealing with group-level effects. Further, ‘winning’, ‘trophy’, ‘work’, ‘scolding’, ‘fight’ etc. reflect some holistic level features which motivate us to study image-level analysis. Similarly, some keywords such as ‘happy’, ‘cheering’, ‘angry’, ‘violence’ etc. tell about the mood of the individuals as well as the group. Thus, the survey motivates us to utilize both image-level features and face-level emotion features of an image. Our experiments are based on the understandings from the survey.

IV Database

To create the database, we have used and extended the images from the GAF 3.0 database [13]. GAF 3.0 has been created via web crawling of various keywords related to social events (for example world cup winners, wedding, family, laughing club, birthday party, siblings, riot, protest and violence etc.). Images are added to GAF 3.0 to get a total of 14,175 images. We split the data into three parts: 9,815 images for training, 4,349 images for validation and 3011 images for testing purposes.

IV-A Data Labeling

The GAF 3.0 database was labelled by 5 annotators (3 females and 2 males) of age group 21-30 years. In order to annotate data, the survey results assist about human perception regarding AGC. We have labelled each image for its cohesiveness in the range [0-3] [2] as shown in Fig. 2. Treadwell et al. [2] argued that it is better to have these four ‘anchor points’ (i.e., strongly agree, agree, disagree and strongly disagree) instead of having low to high scores. The low to high score scaling may vary perception-wise from person to person. Thus, these soft scaled ‘anchor points’ are reliable. Along with GCS, GAF 3.0 database is also labelled with three group emotions (positive, negative and neutral) across the valance axis. Before the annotation, the annotators are familiarized with the concepts of GCS labels [2] with corresponding images.

IV-B Annotation Statistics

We further investigate the agreement between the annotators. The average variance and standard deviation between the annotators are 0.31 and 0.54, respectively. Further, we conduct a principal component analysis on the annotations as shown in Fig. 4. It is evident that approx. 86% of the distribution lies in the first component, which suggests that there is a strong agreement between the annotators. Since the annotations were based on a ‘mutually exclusive category’, we also measure the weighted generalized Cohen’s kappa coefficient [29] to determine the inter-rater agreement. The mean of the kappa coefficients value is 0.51. This also indicates high inter-rater agreeableness.

V Proposed Method

In this section, we discuss our approach for AGC analysis. We examined two networks one of which examines the image as a whole and another examines the facial expression of the group members.

V-A Image-level Analysis

The motivation of this part is to collectively analyze the group and its surroundings. This should also provide contextual information about the group i.e. where the group is and what type of event they are participating in. We use the Inception V3 [30] to train our model for predicting GCS. The main reason behind choosing inception V3 is that it provides a good trade-off between the number of parameters and accuracy in the case of the ImageNet challenge [30]. We have also conducted experiments on several deep Convolutional Neural Networks (CNNs), and results are shown in TABLE I. The inception V3 network is similar to the original work [30], which was proposed for the classification on the ImageNet task except for the last few dense layers including the regression layer. The details of the layers are shown in TABLE II.

In the word cloud of the survey result, people mentioned some group-level emotion-related keywords such as ‘violence’, ‘happy’, ‘angry’, ‘upset’ etc. Thus, we perform experiments with joint training for GCS and group emotion (three classes positive, neutral and negative [13]). The motivation is to explore the usefulness of GCS of a group as an attribute for group emotion prediction. The network structure is the same as shown in TABLE II except for the last layer which predicts three group emotion probabilities and one GCS.

V-B Face-Level Analysis

Motivated by the result of joint training of the AGC and group emotion and survey results (apart from team, group keywords people are mainly focused on the mood like keywords such as angry, cheering, unhappy, violence etc.), we attempt to infer GCS based on the face-level emotion information as well. For facial emotion analysis, we use the recently proposed CapsNet [32] architecture as shown in Fig. 6. In order to overcome the drawbacks of traditional CNNs, Sabour et al. [32] proposed a new CNN like architecture Capsule Network (CapsNet), which keeps the spatial orientation related information along with deep features. Here, capsules are a group of neurons which include the instantiation parameters of a certain object. For example, a face has eyes, nose, lips with certain constraints. The main difference between a CNN and a capsule Network is that the later stores the state of the feature (neuron output) in the form of a vector instead of a scalar. Another salient property of CapsNet is routing by agreement, which means activated capsules follow a hierarchy. Higher level capsules become activated if and only if lower level capsule outputs agree with it. As per [32], CapsNet is invariant to rotation and it can model a spatial hierarchy via dynamic routing and reconstruction regularization. Thus, the network can learn the pattern of viewpoint invariance between the object part and the whole object. From TABLE III, we can observe that CapsNet performs better than the other state-of-the-art networks. CapsNet can explicitly model the pose and illumination of an object. Inspired by this argument, we choose to train CapsNet.

We slightly modified the proposed architecture of CapsNet [32] used for digit classification. CapsNet takes cropped face as input and predicts the seven basic emotions (i.e. happy, neutral, sad, angry, surprise, disgust and fear) as a output. Thus, we get emotion probability predictions for each of the faces present in a group image. Further, we pool the predicted emotion labels by computing the average, maximum and minimum (get $batchsize\times 3\times 7$ dimensional output). This small feature is then fed to two dense layers of 16 and 32 nodes (empirically) respectively before predicting the GCS. This whole network structure is shown in TABLE IV.

V-C Effect of Background

Further, we also investigate how the background effects AGC. We use the segmentation technique to crop people from group images via Deeplab V3plus [31] library. We consider an area-wise threshold, i.e. if the segmented area is less than 50% of the total area of the image then this image is considered for analysis (pipeline is shown in Fig. 6). Thus, we observed that when we use the segmented image for training, then there is a drop (around 0.103 MSE decreased) in performance. It indicates that the background around a person also plays a vital role in the perception of a group’s cohesiveness. The background may reflect something about the social event in which the group is participating and is important for the prediction.

VI Experimental Details and Results

In this section, we discuss the experimental settings and results. First of all, we treat cohesion as a regression problem (as also defined in [2]) and the group emotion as a classification problem (defined in [13]). We use the Keras [33] deep learning library for the implementation.

VI-A Image-Level Analysis Results

We train Inception V3 network for image-level analysis. We initialize the network with ImageNet pre-trained weights and fine-tune the network with SGD optimizer having a learning rate of 0.001 and momentum 0.9 without any learning rate decay. Our image-level experimental results are shown in TABLE VI. With similar hyperparameters, we jointly train an inception V3 network for both emotion and cohesion prediction. The results (TABLE V) show an interesting pattern. When the inception V3 is individually used for group emotion and cohesion prediction, its performance is lower than the joint training. Thus, it suggests that the network learns more relevant representations of group emotion. We can conclude that the emotion and cohesion at group-level are interrelated terms. Human perception behind group emotion and cohesion has some sort of similarity.

This is in accord with psychology studies [14]. It is also interesting to note that the effect of joint training is opposite to the GCS prediction as the prediction error increases. One possible reason is that GCS and group emotion features contradict each other. Let us consider the example of a sobbing family, which has high GCS and negative group emotion and compare that with of a celebrating sports team, which will also have high GCS. In the later, the group emotion will be positive. Scenarios like this may lead to ambiguity during the joint training from the GCS prediction perspective.

VI-B Face-Level Analysis Results

In order to predict GCS, we pre-trained a CapsNet (Capsule Net) on RAF-DB [34]. RAF-DB [34] is a facial expression database containing around 30K single person face images which are labeled for seven basic emotion classes as well as 12 compound emotions. We take basic emotions (i.e. happy, neutral, sad, angry, surprise, disgust and fear) to train our CapsNet. From group images, we extracted faces via MTCNN [35]. After training on RAF-DB, we take the output probability vector for each face in the group image. Further, we compute statistics over these emotion probabilities and pass it through two more dense layers before final cohesion score prediction. Our statistics include maximum, average and minimum of respective emotion probabilities. The motivation behind this is that we need to conclude over a group. Hence, these three intensity level analyses perform better for group-level tasks.

We train a CapsNet with hyperparameters from the original paper (Adam optimizer with default settings in the Keras library and learning rate decay of 0.001 in every 10th epoch to avoid local minima). We train the rest of the network via SGD optimizer with learning rate 0.01 and without any learning rate decay. TABLE VI presents the results of the image-level and face-level networks in GAF Cohesion database. TABLE VII describes 5-fold cross-validation results of the GAF cohesion database. In TABLE VIII, we predict group emotion, when AGC information is used for joint training. The results of the GAF 2.0 show better performance than the other state-of-the-art methods. This shows that cohesiveness information is useful for group emotion prediction.

VII Visualization (Saliency vs Class Activation)

In this section, we discuss visual attributes that our network learns. We visualize the class activation map and discuss its comparison with the saliency. From Fig. 7, we can observe that in spite of non-frontal, rotated, occlusion, blurred faces, CapsNet can handle each case efficiently. Especially, it deals with the rotation and scaling of different objects in an image individually and shows better performance over both occluded and partially occluded images which is beneficial for our problem. Moreover, it did not require data augmentation while training and thus it is efficient regarding time complexity. Similarly, for image-level analysis, (as shown in Fig. 8) the top row activates the background, the second row activates the foreground, the third row activates the subject and the last row activates both the front person and background. In the case of the top row, it activates the background, as the group takes up a small space as compared to the visible background. Similarly, in the second case the foreground is more dominant as compared to the background. In the third row, the main features of the protests that are activated are the banners. In the last picture, it activates both foreground and background, especially the facial region.

The image saliency and class activation have a significant difference. From Fig. 7 and Fig. 8, respectively, it is visible that AGC is not directly labeled as well as predicted via image saliency. Although, there are some regions in images, which are common in both cases. Of course, sometimes the human mind is also influenced by some salient feature of the images.

VIII Conclusion, Limitation and Future Work

The main motivation behind our approach is to achieve human-level perception regarding cohesion and to study the reason behind it. Here, we use the term human-level because we consider human annotation as the ground truth. From our experimental results, it can be deduced that AGC and emotion are interrelated. This work leverages the usefulness of both facial and scene features. We also observed that the newly proposed CapsNet [32] also performs well on facial expression recognition without data augmentation. Although the faces in a group image vary largely, i.e. the face can be occluded, blurred or non-frontal and others. Via visualization, we observed that scene information encodes the background, clothes and various objects in an image. This information is also known as the top-down contextual information. The main limitation of our work is the cultural influence in data annotation as it is related to the perception of cohesion. A potential future direction for our work is to investigate how facial attributes effect AGC prediction. It will be interesting to analyze the role of the body pose of the group members along with the face. Although, the image-level network does encode the body pose, however, its complete contribution requires further investigation. It will be of interest to analyze the fashion quotient of the group by parsing the clothes for patterns and themes, which correspond to specific social events, although, some patterns are already encoded in our scene level analysis. Furthermore, another possible direction is to include kinship related information in the network because irrespective of visual expression, sometimes kinship indicates strong cohesion.

Acknowledgement

We acknowledge the support of NVIDIA for providing us TITAN Xp G5X GPU for research purposes.