The increasing availability of traffic videos functioning on a 24/7/365 time scale has the great potential of increasing the spatio-temporal coverage of traffic accidents, which will help improve traffic safety. However, analyzing footage from hundreds, if not thousands, of traffic cameras in a 24/7/365 working protocol remains an extremely challenging task, as current vision-based approaches primarily focus on extracting raw information, such as vehicle trajectories or individual object detection, but require laborious post-processing to derive actionable insights. We propose SeeUnsafe, a new framework that integrates Multimodal Large Language Model (MLLM) agents to transform video-based traffic accident analysis from a traditional extraction-then-explanation workflow to a more interactive, conversational approach. This shift significantly enhances processing throughput by automating complex tasks like video classification and visual grounding, while improving adaptability by enabling seamless adjustments to diverse traffic scenarios and user-defined queries. Our framework employs a severity-based aggregation strategy to handle videos of various lengths and a novel multimodal prompt to generate structured responses for review and evaluation and enable fine-grained visual grounding. We introduce IMS (Information Matching Score), a new MLLM-based metric for aligning structured responses with ground truth. We conduct extensive experiments on the Toyota Woven Traffic Safety dataset, demonstrating that SeeUnsafe effectively performs accident-aware video classification and visual grounding by leveraging off-the-shelf MLLMs. Source code will be available at \url{https://github.com/ai4ce/SeeUnsafe}.
In the current era of Machine Learning, Transformers have become the de facto approach across a variety of domains, such as computer vision and natural language processing. Transformer-based solutions are the backbone of current state-of-the-art methods for language generation, image and video classification, segmentation, action and object recognition, among many others. Interestingly enough, while these state-of-the-art methods produce impressive results in their respective domains, the problem of understanding the relationship between vision and language is still beyond our reach. In this work, we propose a common ground between vision and language based on events in space and time in an explainable and programmatic way, to connect learning-based vision and language state of the art models and provide a solution to the long standing problem of describing videos in natural language. We validate that our algorithmic approach is able to generate coherent, rich and relevant textual descriptions on videos collected from a variety of datasets, using both standard metrics (e.g. Bleu, ROUGE) and the modern LLM-as-a-Jury approach.
We empirically study autoregressive pre-training from videos. To perform our study, we construct a series of autoregressive video models, called Toto. We treat videos as sequences of visual tokens and train transformer models to autoregressively predict future tokens. Our models are pre-trained on a diverse dataset of videos and images comprising over 1 trillion visual tokens. We explore different architectural, training, and inference design choices. We evaluate the learned visual representations on a range of downstream tasks including image recognition, video classification, object tracking, and robotics. Our results demonstrate that, despite minimal inductive biases, autoregressive pre-training leads to competitive performance across all benchmarks. Finally, we find that scaling our video models results in similar scaling curves to those seen in language models, albeit with a different rate. More details at https://brjathu.github.io/toto/
Challenges in cross-learning involve inhomogeneous or even inadequate amount of training data and lack of resources for retraining large pretrained models. Inspired by transfer learning techniques in NLP, adapters and prefix tuning, this paper presents a new model-agnostic plugin architecture for cross-learning, called CM3T, that adapts transformer-based models to new or missing information. We introduce two adapter blocks: multi-head vision adapters for transfer learning and cross-attention adapters for multimodal learning. Training becomes substantially efficient as the backbone and other plugins do not need to be finetuned along with these additions. Comparative and ablation studies on three datasets Epic-Kitchens-100, MPIIGroupInteraction and UDIVA v0.5 show efficacy of this framework on different recording settings and tasks. With only 12.8% trainable parameters compared to the backbone to process video input and only 22.3% trainable parameters for two additional modalities, we achieve comparable and even better results than the state-of-the-art. CM3T has no specific requirements for training or pretraining and is a step towards bridging the gap between a general model and specific practical applications of video classification.
Automated viewpoint classification in echocardiograms can help under-resourced clinics and hospitals in providing faster diagnosis and screening when expert technicians may not be available. We propose a novel approach towards echocardiographic viewpoint classification. We show that treating viewpoint classification as video classification rather than image classification yields advantage. We propose a CNN-GRU architecture with a novel temporal feature weaving method, which leverages both spatial and temporal information to yield a 4.33\% increase in accuracy over baseline image classification while using only four consecutive frames. The proposed approach incurs minimal computational overhead. Additionally, we publish the Neonatal Echocardiogram Dataset (NED), a professionally-annotated dataset providing sixteen viewpoints and associated echocardipgraphy videos to encourage future work and development in this field. Code available at: https://github.com/satchelfrench/NED
This work aims to predict the popularity of short videos using the videos themselves and their related features. Popularity is measured by four key engagement metrics: view count, like count, comment count, and share count. This study employs video classification models with different architectures and training methods as backbone networks to extract video modality features. Meanwhile, the cleaned video captions are incorporated into a carefully designed prompt framework, along with the video, as input for video-to-text generation models, which generate detailed text-based video content understanding. These texts are then encoded into vectors using a pre-trained BERT model. Based on the six sets of vectors mentioned above, a neural network is trained for each of the four prediction metrics. Moreover, the study conducts data mining and feature engineering based on the video and tabular data, constructing practical features such as the total frequency of hashtag appearances, the total frequency of mention appearances, video duration, frame count, frame rate, and total time online. Multiple machine learning models are trained, and the most stable model, XGBoost, is selected. Finally, the predictions from the neural network and XGBoost models are averaged to obtain the final result.
Both few-shot learning and domain adaptation sub-fields in Computer Vision have seen significant recent progress in terms of the availability of state-of-the-art algorithms and datasets. Frameworks have been developed for each sub-field; however, building a common system or framework that combines both is something that has not been explored. As part of our research, we present the first unified framework that combines domain adaptation for the few-shot learning setting across 3 different tasks - image classification, object detection and video classification. Our framework is highly modular with the capability to support few-shot learning with/without the inclusion of domain adaptation depending on the algorithm. Furthermore, the most important configurable feature of our framework is the on-the-fly setup for incremental $n$-shot tasks with the optional capability to configure the system to scale to a traditional many-shot task. With more focus on Self-Supervised Learning (SSL) for current few-shot learning approaches, our system also supports multiple SSL pre-training configurations. To test our framework's capabilities, we provide benchmarks on a wide range of algorithms and datasets across different task and problem settings. The code is open source has been made publicly available here: https://gitlab.kitware.com/darpa_learn/learn
Audio-visual Zero-Shot Learning (ZSL) has attracted significant attention for its ability to identify unseen classes and perform well in video classification tasks. However, modal imbalance in (G)ZSL leads to over-reliance on the optimal modality, reducing discriminative capabilities for unseen classes. Some studies have attempted to address this issue by modifying parameter gradients, but two challenges still remain: (a) Quality discrepancies, where modalities offer differing quantities and qualities of information for the same concept. (b) Content discrepancies, where sample contributions within a modality vary significantly. To address these challenges, we propose a Discrepancy-Aware Attention Network (DAAN) for Enhanced Audio-Visual ZSL. Our approach introduces a Quality-Discrepancy Mitigation Attention (QDMA) unit to minimize redundant information in the high-quality modality and a Contrastive Sample-level Gradient Modulation (CSGM) block to adjust gradient magnitudes and balance content discrepancies. We quantify modality contributions by integrating optimization and convergence rate for more precise gradient modulation in CSGM. Experiments demonstrates DAAN achieves state-of-the-art performance on benchmark datasets, with ablation studies validating the effectiveness of individual modules.
Human perception integrates multiple modalities, such as vision, hearing, and language, into a unified understanding of the surrounding reality. While recent multimodal models have achieved significant progress by aligning pairs of modalities via contrastive learning, their solutions are unsuitable when scaling to multiple modalities. These models typically align each modality to a designated anchor without ensuring the alignment of all modalities with each other, leading to suboptimal performance in tasks requiring a joint understanding of multiple modalities. In this paper, we structurally rethink the pairwise conventional approach to multimodal learning and we present the novel Gramian Representation Alignment Measure (GRAM), which overcomes the above-mentioned limitations. GRAM learns and then aligns $n$ modalities directly in the higher-dimensional space in which modality embeddings lie by minimizing the Gramian volume of the $k$-dimensional parallelotope spanned by the modality vectors, ensuring the geometric alignment of all modalities simultaneously. GRAM can replace cosine similarity in any downstream method, holding for 2 to $n$ modality and providing more meaningful alignment with respect to previous similarity measures. The novel GRAM-based contrastive loss function enhances the alignment of multimodal models in the higher-dimensional embedding space, leading to new state-of-the-art performance in downstream tasks such as video-audio-text retrieval and audio-video classification. The project page, the code, and the pretrained models are available at https://ispamm.github.io/GRAM/.
Detecting mixed-critical events through computer vision is challenging due to the need for contextual understanding to assess event criticality accurately. Mixed critical events, such as fires of varying severity or traffic incidents, demand adaptable systems that can interpret context to trigger appropriate responses. This paper addresses these challenges by proposing a versatile detection system for smart city applications, offering a solution tested across traffic and fire detection scenarios. Our contributions include an analysis of detection requirements and the development of a system adaptable to diverse applications, advancing automated surveillance for smart cities.
The lack of occlusion data in commonly used action recognition video datasets limits model robustness and impedes sustained performance improvements. We construct OccludeNet, a large-scale occluded video dataset that includes both real-world and synthetic occlusion scene videos under various natural environments. OccludeNet features dynamic tracking occlusion, static scene occlusion, and multi-view interactive occlusion, addressing existing gaps in data. Our analysis reveals that occlusion impacts action classes differently, with actions involving low scene relevance and partial body visibility experiencing greater accuracy degradation. To overcome the limitations of current occlusion-focused approaches, we propose a structural causal model for occluded scenes and introduce the Causal Action Recognition (CAR) framework, which employs backdoor adjustment and counterfactual reasoning. This framework enhances key actor information, improving model robustness to occlusion. We anticipate that the challenges posed by OccludeNet will stimulate further exploration of causal relations in occlusion scenarios and encourage a reevaluation of class correlations, ultimately promoting sustainable performance improvements. The code and full dataset will be released soon.
The advancements in large language models (LLMs) have propelled the improvement of video understanding tasks by incorporating LLMs with visual models. However, most existing LLM-based models (e.g., VideoLLaMA, VideoChat) are constrained to processing short-duration videos. Recent attempts to understand long-term videos by extracting and compressing visual features into a fixed memory size. Nevertheless, those methods leverage only visual modality to merge video tokens and overlook the correlation between visual and textual queries, leading to difficulties in effectively handling complex question-answering tasks. To address the challenges of long videos and complex prompts, we propose AdaCM$^2$, which, for the first time, introduces an adaptive cross-modality memory reduction approach to video-text alignment in an auto-regressive manner on video streams. Our extensive experiments on various video understanding tasks, such as video captioning, video question answering, and video classification, demonstrate that AdaCM$^2$ achieves state-of-the-art performance across multiple datasets while significantly reducing memory usage. Notably, it achieves a 4.5% improvement across multiple tasks in the LVU dataset with a GPU memory consumption reduction of up to 65%.
We present Attend-Fusion, a novel and efficient approach for audio-visual fusion in video classification tasks. Our method addresses the challenge of exploiting both audio and visual modalities while maintaining a compact model architecture. Through extensive experiments on the YouTube-8M dataset, we demonstrate that our Attend-Fusion achieves competitive performance with significantly reduced model complexity compared to larger baseline models.
As violent crimes continue to happen, it becomes necessary to have security cameras that can rapidly identify moments of violence with excellent accuracy. The purpose of this study is to identify how many frames should be analyzed at a time in order to optimize a violence detection model's accuracy as a parameter of the depth of a 3D convolutional network. Previous violence classification models have been created, but their application to live footage may be flawed. In this project, a convolutional neural network was created to analyze optical flow frames of each video. The number of frames analyzed at a time would vary with one, two, three, ten, and twenty frames, and each model would be trained for 20 epochs. The greatest validation accuracy was 94.87% and occurred with the model that analyzed three frames at a time. This means that machine learning models to detect violence may function better when analyzing three frames at a time for this dataset. The methodology used to identify the optimal number of frames to analyze at a time could be used in other applications of video classification, especially those of complex or abstract actions, such as violence.
Deep learning models, in particular \textit{image} models, have recently gained generalisability and robustness. %are becoming more general and robust by the day. In this work, we propose to exploit such advances in the realm of \textit{video} classification. Video foundation models suffer from the requirement of extensive pretraining and a large training time. Towards mitigating such limitations, we propose "\textit{Attention Map (AM) Flow}" for image models, a method for identifying pixels relevant to motion in each input video frame. In this context, we propose two methods to compute AM flow, depending on camera motion. AM flow allows the separation of spatial and temporal processing, while providing improved results over combined spatio-temporal processing (as in video models). Adapters, one of the popular techniques in parameter efficient transfer learning, facilitate the incorporation of AM flow into pretrained image models, mitigating the need for full-finetuning. We extend adapters to "\textit{temporal processing adapters}" by incorporating a temporal processing unit into the adapters. Our work achieves faster convergence, therefore reducing the number of epochs needed for training. Moreover, we endow an image model with the ability to achieve state-of-the-art results on popular action recognition datasets. This reduces training time and simplifies pretraining. We present experiments on Kinetics-400, Something-Something v2, and Toyota Smarthome datasets, showcasing state-of-the-art or comparable results.
As the scale of data and models for video understanding rapidly expand, handling long-form video input in transformer-based models presents a practical challenge. Rather than resorting to input sampling or token dropping, which may result in information loss, token merging shows promising results when used in collaboration with transformers. However, the application of token merging for long-form video processing is not trivial. We begin with the premise that token merging should not rely solely on the similarity of video tokens; the saliency of tokens should also be considered. To address this, we explore various video token merging strategies for long-form video classification, starting with a simple extension of image token merging, moving to region-concentrated merging, and finally proposing a learnable video token merging (VTM) algorithm that dynamically merges tokens based on their saliency. Extensive experimental results show that we achieve better or comparable performances on the LVU, COIN, and Breakfast datasets. Moreover, our approach significantly reduces memory costs by 84% and boosts throughput by approximately 6.89 times compared to baseline algorithms.
Low employment rates in Latin America have contributed to a substantial rise in crime, prompting the emergence of new criminal tactics. For instance, "express robbery" has become a common crime committed by armed thieves, in which they drive motorcycles and assault people in public in a matter of seconds. Recent research has approached the problem by embedding weapon detectors in surveillance cameras; however, these systems are prone to false positives if no counterpart confirms the event. In light of this, we present a distributed IoT system that integrates a computer vision pipeline and object detection capabilities into multiple end-devices, constantly monitoring for the presence of firearms and sharp weapons. Once a weapon is detected, the end-device sends a series of frames to a cloud server that implements a 3DCNN to classify the scene as either a robbery or a normal situation, thus minimizing false positives. The deep learning process to train and deploy weapon detection models uses a custom dataset with 16,799 images of firearms and sharp weapons. The best-performing model, YOLOv5s, optimized using TensorRT, achieved a final mAP of 0.87 running at 4.43 FPS. Additionally, the 3DCNN demonstrated 0.88 accuracy in detecting abnormal situations. Extensive experiments validate that the proposed system significantly reduces false positives while autonomously monitoring multiple locations in real-time.
Dense features, customized for different business scenarios, are essential in short video classification. However, their complexity, specific adaptation requirements, and high computational costs make them resource-intensive and less accessible during online inference. Consequently, these dense features are categorized as `Privileged Dense Features'.Meanwhile, end-to-end multi-modal models have shown promising results in numerous computer vision tasks. In industrial applications, prioritizing end-to-end multi-modal features, can enhance efficiency but often leads to the loss of valuable information from historical privileged dense features. To integrate both features while maintaining efficiency and manageable resource costs, we present Confidence-aware Privileged Feature Distillation (CPFD), which empowers features of an end-to-end multi-modal model by adaptively distilling privileged features during training. Unlike existing privileged feature distillation (PFD) methods, which apply uniform weights to all instances during distillation, potentially causing unstable performance across different business scenarios and a notable performance gap between teacher model (Dense Feature enhanced multimodal-model DF-X-VLM) and student model (multimodal-model only X-VLM), our CPFD leverages confidence scores derived from the teacher model to adaptively mitigate the performance variance with the student model. We conducted extensive offline experiments on five diverse tasks demonstrating that CPFD improves the video classification F1 score by 6.76% compared with end-to-end multimodal-model (X-VLM) and by 2.31% with vanilla PFD on-average. And it reduces the performance gap by 84.6% and achieves results comparable to teacher model DF-X-VLM. The effectiveness of CPFD is further substantiated by online experiments, and our framework has been deployed in production systems for over a dozen models.
The rise of short-form videos on platforms like TikTok has brought new challenges in safeguarding young viewers from inappropriate content. Traditional moderation methods often fall short in handling the vast and rapidly changing landscape of user-generated videos, increasing the risk of children encountering harmful material. This paper introduces TikGuard, a transformer-based deep learning approach aimed at detecting and flagging content unsuitable for children on TikTok. By using a specially curated dataset, TikHarm, and leveraging advanced video classification techniques, TikGuard achieves an accuracy of 86.7%, showing a notable improvement over existing methods in similar contexts. While direct comparisons are limited by the uniqueness of the TikHarm dataset, TikGuard's performance highlights its potential in enhancing content moderation, contributing to a safer online experience for minors. This study underscores the effectiveness of transformer models in video classification and sets a foundation for future research in this area.
Edge computing's growing prominence, due to its ability to reduce communication latency and enable real-time processing, is promoting the rise of high-performance, heterogeneous System-on-Chip solutions. While current approaches often involve scaling down modern hardware, the performance characteristics of neural network workloads on these platforms can vary significantly, especially when it comes to parallel processing, which is a critical consideration for edge deployments. To address this, we conduct a comprehensive study comparing the latency and throughput of various linear algebra and neural network inference tasks across CPU-only, CPU/GPU, and CPU/NPU integrated solutions. {We find that the Neural Processing Unit (NPU) excels in matrix-vector multiplication (58.6% faster) and some neural network tasks (3.2$\times$ faster for video classification and large language models). GPU outperforms in matrix multiplication (22.6% faster) and LSTM networks (2.7$\times$ faster) while CPU excels at less parallel operations like dot product. NPU-based inference offers a balance of latency and throughput at lower power consumption. GPU-based inference, though more energy-intensive, performs best with large dimensions and batch sizes. We highlight the potential of heterogeneous computing solutions for edge AI, where diverse compute units can be strategically leveraged to boost accurate and real-time inference.