Onur Cezmi Mutlu

Yang Qian, Ali Kargarandehkordi, Onur Cezmi Mutlu et al.

Emotions play an essential role in human communication. Developing computer vision models for automatic recognition of emotion expression can aid in a variety of domains, including robotics, digital behavioral healthcare, and media analytics. There are three types of emotional representations which are traditionally modeled in affective computing research: Action Units, Valence Arousal (VA), and Categorical Emotions. As part of an effort to move beyond these representations towards more fine-grained labels, we describe our submission to the newly introduced Emotional Reaction Intensity (ERI) Estimation challenge in the 5th competition for Affective Behavior Analysis in-the-Wild (ABAW). We developed four deep neural networks trained in the visual domain and a multimodal model trained with both visual and audio features to predict emotion reaction intensity. Our best performing model on the Hume-Reaction dataset achieved an average Pearson correlation coefficient of 0.4080 on the test set using a pre-trained ResNet50 model. This work provides a first step towards the development of production-grade models which predict emotion reaction intensities rather than discrete emotion categories.

Marie Huynh, Aaron Kline, Saimourya Surabhi et al.

Early detection of autism, a neurodevelopmental disorder marked by social communication challenges, is crucial for timely intervention. Recent advancements have utilized naturalistic home videos captured via the mobile application GuessWhat. Through interactive games played between children and their guardians, GuessWhat has amassed over 3,000 structured videos from 382 children, both diagnosed with and without Autism Spectrum Disorder (ASD). This collection provides a robust dataset for training computer vision models to detect ASD-related phenotypic markers, including variations in emotional expression, eye contact, and head movements. We have developed a protocol to curate high-quality videos from this dataset, forming a comprehensive training set. Utilizing this set, we trained individual LSTM-based models using eye gaze, head positions, and facial landmarks as input features, achieving test AUCs of 86%, 67%, and 78%, respectively. To boost diagnostic accuracy, we applied late fusion techniques to create ensemble models, improving the overall AUC to 90%. This approach also yielded more equitable results across different genders and age groups. Our methodology offers a significant step forward in the early detection of ASD by potentially reducing the reliance on subjective assessments and making early identification more accessibly and equitable.

Onur Cezmi Mutlu, Mohammadmahdi Honarmand, Saimourya Surabhi et al.

We introduce Temporal consistency for Test-time adaptation (TempT) a novel method for test-time adaptation on videos through the use of temporal coherence of predictions across sequential frames as a self-supervision signal. TempT is an approach with broad potential applications in computer vision tasks including facial expression recognition (FER) in videos. We evaluate TempT performance on the AffWild2 dataset. Our approach focuses solely on the unimodal visual aspect of the data and utilizes a popular 2D CNN backbone in contrast to larger sequential or attention-based models used in other approaches. Our preliminary experimental results demonstrate that TempT has competitive performance compared to the previous years reported performances and its efficacy provides a compelling proof-of-concept for its use in various real-world applications.

Yang Qian, Yinan Sun, Ali Kargarandehkordi et al.

The increasing variety and quantity of tagged multimedia content on a variety of online platforms offer a unique opportunity to advance the field of human action recognition. In this study, we utilize 283,582 unique, unlabeled TikTok video clips, categorized into 386 hashtags, to train a domain-specific foundation model for action recognition. We employ VideoMAE V2, an advanced model integrating Masked Autoencoders (MAE) with Vision Transformers (ViT), pre-trained on this diverse collection of unstructured videos. Our model, fine-tuned on established action recognition benchmarks such as UCF101 and HMDB51, achieves state-of-the-art results: 99.05% on UCF101, 86.08% on HMDB51, 85.51% on Kinetics-400, and 74.27% on Something-Something V2 using the ViT-giant backbone. These results highlight the potential of using unstructured and unlabeled videos as a valuable source of diverse and dynamic content for training foundation models. Our investigation confirms that while initial increases in pre-training data volume significantly enhance model performance, the gains diminish as the dataset size continues to expand. Our findings emphasize two critical axioms in self-supervised learning for computer vision: (1) additional pre-training data can yield diminishing benefits for some datasets and (2) quality is more important than quantity in self-supervised learning, especially when building foundation models.

Mohammadmahdi Honarmand, Onur Cezmi Mutlu, Parnian Azizian et al.

Robust facial expression recognition in unconstrained, "in-the-wild" environments remains challenging due to significant domain shifts between training and testing distributions. Test-time adaptation (TTA) offers a promising solution by adapting pre-trained models during inference without requiring labeled test data. However, existing TTA approaches typically rely on manually selecting which parameters to update, potentially leading to suboptimal adaptation and high computational costs. This paper introduces a novel Fisher-driven selective adaptation framework that dynamically identifies and updates only the most critical model parameters based on their importance as quantified by Fisher information. By integrating this principled parameter selection approach with temporal consistency constraints, our method enables efficient and effective adaptation specifically tailored for video-based facial expression recognition. Experiments on the challenging AffWild2 benchmark demonstrate that our approach significantly outperforms existing TTA methods, achieving a 7.7% improvement in F1 score over the base model while adapting only 22,000 parameters-more than 20 times fewer than comparable methods. Our ablation studies further reveal that parameter importance can be effectively estimated from minimal data, with sampling just 1-3 frames sufficient for substantial performance gains. The proposed approach not only enhances recognition accuracy but also dramatically reduces computational overhead, making test-time adaptation more practical for real-world affective computing applications.

Anish Lakkapragada, Aaron Kline, Onur Cezmi Mutlu et al.

A formal autism diagnosis can be an inefficient and lengthy process. Families may wait months or longer before receiving a diagnosis for their child despite evidence that earlier intervention leads to better treatment outcomes. Digital technologies which detect the presence of behaviors related to autism can scale access to pediatric diagnoses. This work aims to demonstrate the feasibility of deep learning technologies for detecting hand flapping from unstructured home videos as a first step towards validating whether models and digital technologies can be leveraged to aid with autism diagnoses. We used the Self-Stimulatory Behavior Dataset (SSBD), which contains 75 videos of hand flapping, head banging, and spinning exhibited by children. From all the hand flapping videos, we extracted 100 positive and control videos of hand flapping, each between 2 to 5 seconds in duration. Utilizing both landmark-driven-approaches and MobileNet V2's pretrained convolutional layers, our highest performing model achieved a testing F1 score of 84% (90% precision and 80% recall) when evaluating with 5-fold cross validation 100 times. This work provides the first step towards developing precise deep learning methods for activity detection of autism-related behaviors.

Peter Washington, Aaron Kline, Onur Cezmi Mutlu et al.

Activity recognition computer vision algorithms can be used to detect the presence of autism-related behaviors, including what are termed "restricted and repetitive behaviors", or stimming, by diagnostic instruments. The limited data that exist in this domain are usually recorded with a handheld camera which can be shaky or even moving, posing a challenge for traditional feature representation approaches for activity detection which mistakenly capture the camera's motion as a feature. To address these issues, we first document the advantages and limitations of current feature representation techniques for activity recognition when applied to head banging detection. We then propose a feature representation consisting exclusively of head pose keypoints. We create a computer vision classifier for detecting head banging in home videos using a time-distributed convolutional neural network (CNN) in which a single CNN extracts features from each frame in the input sequence, and these extracted features are fed as input to a long short-term memory (LSTM) network. On the binary task of predicting head banging and no head banging within videos from the Self Stimulatory Behaviour Dataset (SSBD), we reach a mean F1-score of 90.77% using 3-fold cross validation (with individual fold F1-scores of 83.3%, 89.0%, and 100.0%) when ensuring that no child who appeared in the train set was in the test set for all folds. This work documents a successful technique for training a computer vision classifier which can detect human motion with few training examples and even when the camera recording the source clips is unstable. The general methods described here can be applied by designers and developers of interactive systems towards other human motion and pose classification problems used in mobile and ubiquitous interactive systems.

Peter Washington, Onur Cezmi Mutlu, Emilie Leblanc et al.

Emotion classifiers traditionally predict discrete emotions. However, emotion expressions are often subjective, thus requiring a method to handle subjective labels. We explore the use of crowdsourcing to acquire reliable soft-target labels and evaluate an emotion detection classifier trained with these labels. We center our study on the Child Affective Facial Expression (CAFE) dataset, a gold standard collection of images depicting pediatric facial expressions along with 100 human labels per image. To test the feasibility of crowdsourcing to generate these labels, we used Microworkers to acquire labels for 207 CAFE images. We evaluate both unfiltered workers as well as workers selected through a short crowd filtration process. We then train two versions of a classifiers on soft-target CAFE labels using the original 100 annotations provided with the dataset: (1) a classifier trained with traditional one-hot encoded labels, and (2) a classifier trained with vector labels representing the distribution of CAFE annotator responses. We compare the resulting softmax output distributions of the two classifiers with a 2-sample independent t-test of L1 distances between the classifier's output probability distribution and the distribution of human labels. While agreement with CAFE is weak for unfiltered crowd workers, the filtered crowd agree with the CAFE labels 100% of the time for many emotions. While the F1-score for a one-hot encoded classifier is much higher (94.33% vs. 78.68%) with respect to the ground truth CAFE labels, the output probability vector of the crowd-trained classifier more closely resembles the distribution of human labels (t=3.2827, p=0.0014). Reporting an emotion probability distribution that accounts for the subjectivity of human interpretation. Crowdsourcing, including a sufficient filtering mechanism, is a feasible solution for acquiring soft-target labels.

Peter Washington, Haik Kalantarian, John Kent et al.

Background: Automated emotion classification could aid those who struggle to recognize emotions, including children with developmental behavioral conditions such as autism. However, most computer vision emotion recognition models are trained on adult emotion and therefore underperform when applied to child faces. Objective: We designed a strategy to gamify the collection and labeling of child emotion-enriched images to boost the performance of automatic child emotion recognition models to a level closer to what will be needed for digital health care approaches. Methods: We leveraged our prototype therapeutic smartphone game, GuessWhat, which was designed in large part for children with developmental and behavioral conditions, to gamify the secure collection of video data of children expressing a variety of emotions prompted by the game. Independently, we created a secure web interface to gamify the human labeling effort, called HollywoodSquares, tailored for use by any qualified labeler. We gathered and labeled 2155 videos, 39,968 emotion frames, and 106,001 labels on all images. With this drastically expanded pediatric emotion-centric database (>30 times larger than existing public pediatric emotion data sets), we trained a convolutional neural network (CNN) computer vision classifier of happy, sad, surprised, fearful, angry, disgust, and neutral expressions evoked by children. Results: The classifier achieved a 66.9% balanced accuracy and 67.4% F1-score on the entirety of the Child Affective Facial Expression (CAFE) as well as a 79.1% balanced accuracy and 78% F1-score on CAFE Subset A, a subset containing at least 60% human agreement on emotions labels. This performance is at least 10% higher than all previously developed classifiers evaluated against CAFE, the best of which reached a 56% balanced accuracy even when combining "anger" and "disgust" into a single class.