Ranked from Within: Ranking Large Multimodal Models for Visual Question Answering Without Labels

As large multimodal models (LMMs) are increasingly deployed across diverse applications, the need for adaptable, real-world model ranking has become paramount. Traditional evaluation methods are largely dataset-centric, relying on fixed, labeled datasets and supervised metrics, which are resource-intensive and may lack generalizability to novel scenarios, highlighting the importance of unsupervised ranking. In this work, we explore unsupervised model ranking for LMMs by leveraging their uncertainty signals, such as softmax probabilities. We evaluate state-of-the-art LMMs (e.g., LLaVA) across visual question answering benchmarks, analyzing how uncertainty-based metrics can reflect model performance. Our findings show that uncertainty scores derived from softmax distributions provide a robust, consistent basis for ranking models across varied tasks. This finding enables the ranking of LMMs on real-world, unlabeled data for visual question answering, providing a practical approach for selecting models across diverse domains without requiring manual annotation.

When Spatial meets Temporal in Action Recognition

Video action recognition has made significant strides, but challenges remain in effectively using both spatial and temporal information. While existing methods often focus on either spatial features (e.g., object appearance) or temporal dynamics (e.g., motion), they rarely address the need for a comprehensive integration of both. Capturing the rich temporal evolution of video frames, while preserving their spatial details, is crucial for improving accuracy. In this paper, we introduce the Temporal Integration and Motion Enhancement (TIME) layer, a novel preprocessing technique designed to incorporate temporal information. The TIME layer generates new video frames by rearranging the original sequence, preserving temporal order while embedding $N^2$ temporally evolving frames into a single spatial grid of size $N \times N$. This transformation creates new frames that balance both spatial and temporal information, making them compatible with existing video models. When $N=1$, the layer captures rich spatial details, similar to existing methods. As $N$ increases ($N\geq2$), temporal information becomes more prominent, while the spatial information decreases to ensure compatibility with model inputs. We demonstrate the effectiveness of the TIME layer by integrating it into popular action recognition models, such as ResNet-50, Vision Transformer, and Video Masked Autoencoders, for both RGB and depth video data. Our experiments show that the TIME layer enhances recognition accuracy, offering valuable insights for video processing tasks.

Visual Prompting in LLMs for Enhancing Emotion Recognition

Vision Large Language Models (VLLMs) are transforming the intersection of computer vision and natural language processing. Nonetheless, the potential of using visual prompts for emotion recognition in these models remains largely unexplored and untapped. Traditional methods in VLLMs struggle with spatial localization and often discard valuable global context. To address this problem, we propose a Set-of-Vision prompting (SoV) approach that enhances zero-shot emotion recognition by using spatial information, such as bounding boxes and facial landmarks, to mark targets precisely. SoV improves accuracy in face count and emotion categorization while preserving the enriched image context. Through a battery of experimentation and analysis of recent commercial or open-source VLLMs, we evaluate the SoV model's ability to comprehend facial expressions in natural environments. Our findings demonstrate the effectiveness of integrating spatial visual prompts into VLLMs for improving emotion recognition performance.

Toward a Holistic Evaluation of Robustness in CLIP Models

Contrastive Language-Image Pre-training (CLIP) models have shown significant potential, particularly in zero-shot classification across diverse distribution shifts. Building on existing evaluations of overall classification robustness, this work aims to provide a more comprehensive assessment of CLIP by introducing several new perspectives. First, we investigate their robustness to variations in specific visual factors. Second, we assess two critical safety objectives--confidence uncertainty and out-of-distribution detection--beyond mere classification accuracy. Third, we evaluate the finesse with which CLIP models bridge the image and text modalities. Fourth, we extend our examination to 3D awareness in CLIP models, moving beyond traditional 2D image understanding. Finally, we explore the interaction between vision and language encoders within modern large multimodal models (LMMs) that utilize CLIP as the visual backbone, focusing on how this interaction impacts classification robustness. In each aspect, we consider the impact of six factors on CLIP models: model architecture, training distribution, training set size, fine-tuning, contrastive loss, and test-time prompts. Our study uncovers several previously unknown insights into CLIP. For instance, the architecture of the visual encoder in CLIP plays a significant role in their robustness against 3D corruption. CLIP models tend to exhibit a bias towards shape when making predictions. Moreover, this bias tends to diminish after fine-tuning on ImageNet. Vision-language models like LLaVA, leveraging the CLIP vision encoder, could exhibit benefits in classification performance for challenging categories over CLIP alone. Our findings are poised to offer valuable guidance for enhancing the robustness and reliability of CLIP models.

LEGO: Learnable Expansion of Graph Operators for Multi-Modal Feature Fusion

In computer vision tasks, features often come from diverse representations, domains, and modalities, such as text, images, and videos. Effectively fusing these features is essential for robust performance, especially with the availability of powerful pre-trained models like vision-language models. However, common fusion methods, such as concatenation, element-wise operations, and non-linear techniques, often fail to capture structural relationships, deep feature interactions, and suffer from inefficiency or misalignment of features across domains. In this paper, we shift from high-dimensional feature space to a lower-dimensional, interpretable graph space by constructing similarity graphs that encode feature relationships at different levels, e.g., clip, frame, patch, token, etc. To capture deeper interactions, we use graph power expansions and introduce a learnable graph fusion operator to combine these graph powers for more effective fusion. Our approach is relationship-centric, operates in a homogeneous space, and is mathematically principled, resembling element-wise similarity score aggregation via multilinear polynomials. We demonstrate the effectiveness of our graph-based fusion method on video anomaly detection, showing strong performance across multi-representational, multi-modal, and multi-domain feature fusion tasks.

TrackNetV4: Enhancing Fast Sports Object Tracking with Motion Attention Maps

Accurately detecting and tracking high-speed, small objects, such as balls in sports videos, is challenging due to factors like motion blur and occlusion. Although recent deep learning frameworks like TrackNetV1, V2, and V3 have advanced tennis ball and shuttlecock tracking, they often struggle in scenarios with partial occlusion or low visibility. This is primarily because these models rely heavily on visual features without explicitly incorporating motion information, which is crucial for precise tracking and trajectory prediction. In this paper, we introduce an enhancement to the TrackNet family by fusing high-level visual features with learnable motion attention maps through a motion-aware fusion mechanism, effectively emphasizing the moving ball's location and improving tracking performance. Our approach leverages frame differencing maps, modulated by a motion prompt layer, to highlight key motion regions over time. Experimental results on the tennis ball and shuttlecock datasets show that our method enhances the tracking performance of both TrackNetV2 and V3. We refer to our lightweight, plug-and-play solution, built on top of the existing TrackNet, as TrackNetV4.

MRAC Track 1: 2nd Workshop on Multimodal, Generative and Responsible Affective Computing

With the rapid advancements in multimodal generative technology, Affective Computing research has provoked discussion about the potential consequences of AI systems equipped with emotional intelligence. Affective Computing involves the design, evaluation, and implementation of Emotion AI and related technologies aimed at improving people's lives. Designing a computational model in affective computing requires vast amounts of multimodal data, including RGB images, video, audio, text, and physiological signals. Moreover, Affective Computing research is deeply engaged with ethical considerations at various stages-from training emotionally intelligent models on large-scale human data to deploying these models in specific applications. Fundamentally, the development of any AI system must prioritize its impact on humans, aiming to augment and enhance human abilities rather than replace them, while drawing inspiration from human intelligence in a safe and responsible manner. The MRAC 2024 Track 1 workshop seeks to extend these principles from controlled, small-scale lab environments to real-world, large-scale contexts, emphasizing responsible development. The workshop also aims to highlight the potential implications of generative technology, along with the ethical consequences of its use, to researchers and industry professionals. To the best of our knowledge, this is the first workshop series to comprehensively address the full spectrum of multimodal, generative affective computing from a responsible AI perspective, and this is the second iteration of this workshop. Webpage: https://react-ws.github.io/2024/

MIP-GAF: A MLLM-annotated Benchmark for Most Important Person Localization and Group Context Understanding

Estimating the Most Important Person (MIP) in any social event setup is a challenging problem mainly due to contextual complexity and scarcity of labeled data. Moreover, the causality aspects of MIP estimation are quite subjective and diverse. To this end, we aim to address the problem by annotating a large-scale `in-the-wild' dataset for identifying human perceptions about the `Most Important Person (MIP)' in an image. The paper provides a thorough description of our proposed Multimodal Large Language Model (MLLM) based data annotation strategy, and a thorough data quality analysis. Further, we perform a comprehensive benchmarking of the proposed dataset utilizing state-of-the-art MIP localization methods, indicating a significant drop in performance compared to existing datasets. The performance drop shows that the existing MIP localization algorithms must be more robust with respect to `in-the-wild' situations. We believe the proposed dataset will play a vital role in building the next-generation social situation understanding methods. The code and data is available at https://github.com/surbhimadan92/MIP-GAF.

MADE-for-ASD: A Multi-Atlas Deep Ensemble Network for Diagnosing Autism Spectrum Disorder

In response to the global need for efficient early diagnosis of Autism Spectrum Disorder (ASD), this paper bridges the gap between traditional, time-consuming diagnostic methods and potential automated solutions. We propose a multi-atlas deep ensemble network, MADE-for-ASD, that integrates multiple atlases of the brain's functional magnetic resonance imaging (fMRI) data through a weighted deep ensemble network. Our approach integrates demographic information into the prediction workflow, which enhances ASD diagnosis performance and offers a more holistic perspective on patient profiling. We experiment with the well-known publicly available ABIDE (Autism Brain Imaging Data Exchange) I dataset, consisting of resting state fMRI data from 17 different laboratories around the globe. Our proposed system achieves 75.20% accuracy on the entire dataset and 96.40% on a specific subset $-$ both surpassing reported ASD diagnosis accuracy in ABIDE I fMRI studies. Specifically, our model improves by 4.4 percentage points over prior works on the same amount of data. The model exhibits a sensitivity of 82.90% and a specificity of 69.70% on the entire dataset, and 91.00% and 99.50%, respectively, on the specific subset. We leverage the F-score to pinpoint the top 10 ROI in ASD diagnosis, such as \emph{precuneus} and anterior \emph{cingulate/ventromedial}. The proposed system can potentially pave the way for more cost-effective, efficient and scalable strategies in ASD diagnosis. Codes and evaluations are publicly available at TBA.

Motion meets Attention: Video Motion Prompts

Videos contain rich spatio-temporal information. Traditional methods for extracting motion, used in tasks such as action recognition, often rely on visual contents rather than precise motion features. This phenomenon is referred to as 'blind motion extraction' behavior, which proves inefficient in capturing motions of interest due to a lack of motion-guided cues. Recently, attention mechanisms have enhanced many computer vision tasks by effectively highlighting salient visual areas. Inspired by this, we propose using a modified Sigmoid function with learnable slope and shift parameters as an attention mechanism to activate and modulate motion signals derived from frame differencing maps. This approach generates a sequence of attention maps that enhance the processing of motion-related video content. To ensure temporally continuity and smoothness of the attention maps, we apply pair-wise temporal attention variation regularization to remove unwanted motions (e.g., noise) while preserving important ones. We then perform Hadamard product between each pair of attention maps and the original video frames to highlight the evolving motions of interest over time. These highlighted motions, termed video motion prompts, are subsequently used as inputs to the model instead of the original video frames. We formalize this process as a motion prompt layer and incorporate the regularization term into the loss function to learn better motion prompts. This layer serves as an adapter between the model and the video data, bridging the gap between traditional 'blind motion extraction' and the extraction of relevant motions of interest.