arXiv subjects
Publications and source records attributed to Cathal Gurrin.
This report summarizes the contributions and results of the 1st Asynchronous CASTLE Challenge at the Joint Egocentric Vision Workshop in conjunction with CVPR 2026.
Long-horizon video editing agents receive final-product feedback only after many interdependent decisions. Yet editing quality is subjective, admits multiple valid solutions, and is not meaningfully calibrated across heterogeneous requests, making a global scalar objective both ambiguous and temporally uninformative. Our key observation is that fixing the request, materials, and production constraints converts this subjective objective into an ordinal comparison among directly comparable alternatives. We introduce Group-Relative Preference Backpropagation (GRPB), which transforms same-task rankings into zero-sum advantages and redistributes them as bounded credit over semantic editing segments. A lagged allocator and guarded transmission prevent current judgments or unreliable estimates from directly shaping the same rollout group. We manually construct a project-disjoint, horizon-stratified suite of realistic editing tasks for training and controlled evaluation. Across matched baselines, credit interventions, external benchmarking, and blinded human evaluation, GRPB improves both editing behavior and rendered products. The resulting 9B Crayotter model surpasses several proprietary systems on AgenticVBench, supporting task-local preference reduction as a practical approach to learning from subjective, delayed outcomes. Code and all supporting materials are publicly available at https://github.com/idwts/Crayotter.
Language models (LMs) are trained to excel at predicting the next word in the sequence given prior context, and humans also share this predictability in reading comprehension. Neuroscience research reveals that next-word predictability influences brain response, as recorded at millisecond resolution using electroencephalography (EEG). While our evidence indicates that advanced LMs achieve accuracies closely aligned with human performance at the next-word prediction task, this raises the question: Does higher prediction accuracy necessarily mean that these models adequately capture the cognitive signals associated with human reading comprehension? Here, we generate regressors for both humans and LMs based on two information measures, including top-1 prediction and surprisal, to predict event-related potential (ERP) elicited from EEG recordings which reflect different stages of cognitive processing during reading. We argue that modelling ERP patterns offers fine-grained analysis of the cognitive plausibility of various LMs during reading. Our results indicate that only surprisal potentially correlates with language-processing ERPs, especially for open-class words with high semantic content. Moreover, our findings challenge the assumption that scaling LMs with increased parameters and computational budgets will consistently lead to improved convergence with human-like linguistic processing.
Humans invented reading and have passed down this complex skill across generations through language. This study provides empirical evidence of the neural mechanisms underlying bottom-up (related to high-order linguistic structure) and top-down (related to next-word predictability) processes, which interact to guide comprehension during reading. While previous studies have focused on either the N400 effects of predictability or lexical categories, research on how predictability influences N400 responses across different lexical categories is limited, mainly due to constraints in publicly available datasets. Here, we examine how predictability influences brain responses, recorded at millisecond resolution using electroencephalography (EEG), with a focus on the N400 time window (300-500 ms post-stimulus) across different lexical and grammatical categories. Our results indicate that significant differences in N400 responses between high and low cloze probability levels were more pronounced for content words than function words. Among the two primary content categories, verbs exhibited greater N400 differences than nouns, while nouns carried more distinct information about their predictability than verbs. Moreover, we demonstrate that the decoding technique is more effective than the event-related potential (ERP) traditional analysis in capturing more detailed and distinct representations of cognitive processes over time.
Vision-Language Models (VLMs) are typically pre-trained on large-scale image-text datasets to capture semantic correspondences between visual content and natural language. However, they remain surprisingly brittle to negation: models often rely on shallow word co-occurrence and are easily distracted by misleading or irrelevant textual cues, even when their overall retrieval or classification performance is strong. Moreover, directly finetuning on negation data can interfere with previously acquired knowledge, causing noticeable degradation on standard vision-language benchmarks. To tackle these issues, this work introduces HANCLIP (Hyperbolic + Angular + Negation), a family of VLMs that explicitly restructures the embedding space to encode "what an image is not" alongside "what it is." HANCLIP is trained on a compact set of 20,000 image-text quadruplets and combines a hyperbolic formulation, which models hierarchical semantic relations and asymmetries, with an angular triplet objective that drives systematic separation between negated descriptions and their corresponding positives. This geometry-aware design strengthens negation sensitivity while preserving the global structure of pretrained representations, rather than overwriting them. Extensive experiments across multiple vision-language tasks show that HANCLIP delivers consistent gains on the negation-focused NegBench benchmark, while maintaining competitive or improved performance on standard classification and image-text retrieval benchmarks. The framework is model-agnostic and can be plugged into CLIP, LongCLIP, SmartCLIP, and HiMo-CLIP without large-scale retraining, demonstrating that a carefully designed geometric objective can substantially extend the reasoning capabilities of existing VLMs using only modest additional data.
Image-Guided Retrieval with Optional Text (IGROT) unifies visual retrieval (without text) and composed retrieval (with text). Despite its relevance in applications like Google Image and Bing, progress has been limited by the lack of an accessible benchmark and methods that balance performance across subtasks. Large-scale datasets such as MagicLens are comprehensive but computationally prohibitive, while existing models often favor either visual or compositional queries. We introduce FIGROTD, a lightweight yet high-quality IGROT dataset with 16,474 training triplets and 1,262 test triplets across CIR, SBIR, and CSTBIR. To reduce redundancy, we propose the Variance Guided Feature Mask (VaGFeM), which selectively enhances discriminative dimensions based on variance statistics. We further adopt a dual-loss design (InfoNCE + Triplet) to improve compositional reasoning. Trained on FIGROTD, VaGFeM achieves competitive results on nine benchmarks, reaching 34.8 mAP@10 on CIRCO and 75.7 mAP@200 on Sketchy, outperforming stronger baselines despite fewer triplets.
Image-Guided Retrieval with Optional Text (IGROT) is a general retrieval setting where a query consists of an anchor image, with or without accompanying text, aiming to retrieve semantically relevant target images. This formulation unifies two major tasks: Composed Image Retrieval (CIR) and Sketch-Based Image Retrieval (SBIR). In this work, we address IGROT under low-data supervision by introducing UNION, a lightweight and generalisable target representation that fuses the image embedding with a null-text prompt. Unlike traditional approaches that rely on fixed target features, UNION enhances semantic alignment with multimodal queries while requiring no architectural modifications to pretrained vision-language models. With only 5,000 training samples - from LlavaSCo for CIR and Training-Sketchy for SBIR - our method achieves competitive results across benchmarks, including CIRCO mAP@50 of 38.5 and Sketchy mAP@200 of 82.7, surpassing many heavily supervised baselines. This demonstrates the robustness and efficiency of UNION in bridging vision and language across diverse query types.
This report presents the results of the 14th Video Browser Showdown, held at the 2025 International Conference on Multimedia Modeling on the 8th of January 2025 in Nara, Japan.
We introduce OpenLifelogQA, a large-scale open-ended lifelog QA dataset constructed from 18 months of multimodal lifelog data. Lifelogging is the passive collection and analysis of personal daily activities using wearable devices, producing rich multimodal data such as images, locations, and biometrics. Question answering (QA) over lifelog data enables users to interactively query their own experiences, supporting applications in memory support, lifestyle analysis, and personal assistance. OpenLifelogQA contains 14,187 Q&A pairs spanning multiple question types and difficulty levels, designed to support robust evaluation in realistic settings. Compared with prior resources, OpenLifelogQA offers greater diversity and practicality for real-world applications. To establish baselines, we evaluate the LLaVA-NeXT-Interleave 7B model, achieving 89.7% BERTScore, 25.87% ROUGE-L, and an average LLM Score of 3.97. By releasing OpenLifelogQA, we aim to promote future research on lifelog technologies, paving the way for personal lifelog assistants capable of memory augmentation, healthcare support, and lifestyle coaching.
This paper addresses two main objectives. Firstly, we demonstrate the impressive performance of the LLaVA-NeXT-interleave on 22 datasets across three different tasks: Multi-Image Reasoning, Documents and Knowledge-Based Understanding and Interactive Multi-Modal Communication. Secondly, we add the Dense Channel Integration (DCI) connector to the LLaVA-NeXT-Interleave and compare its performance against the standard model. We find that the standard model achieves the highest overall accuracy, excelling in vision-heavy tasks like VISION, NLVR2, and Fashion200K. Meanwhile, the DCI-enhanced version shows particular strength on datasets requiring deeper semantic coherence or structured change understanding such as MIT-States_PropertyCoherence and SlideVQA. Our results highlight the potential of combining powerful foundation models with plug-and-play techniques for Interleave tasks. The code is available at https://github.com/dinhvietcuong1996/icme25-inova.
The ACM Lifelog Search Challenge (LSC) is a venue that welcomes and compares systems that support the exploration of lifelog data, and in particular the retrieval of specific information, through an interactive competition format. This paper reviews the recent advances in interactive lifelog retrieval as demonstrated at the ACM LSC from 2022 to 2024. Through a detailed comparative analysis, we highlight key improvements across three main retrieval tasks: known-item search, question answering, and ad-hoc search. Our analysis identifies trends such as the widespread adoption of embedding-based retrieval methods (e.g., CLIP, BLIP), increased integration of large language models (LLMs) for conversational retrieval, and continued innovation in multimodal and collaborative search interfaces. We further discuss how specific retrieval techniques and user interface (UI) designs have impacted system performance, emphasizing the importance of balancing retrieval complexity with usability. Our findings indicate that embedding-driven approaches combined with LLMs show promise for lifelog retrieval systems. Likewise, improving UI design can enhance usability and efficiency. Additionally, we recommend reconsidering multi-instance system evaluations within the expert track to better manage variability in user familiarity and configuration effectiveness.
Lifelogging involves continuously capturing personal data through wearable cameras, providing an egocentric view of daily activities. Lifelog retrieval aims to search and retrieve relevant moments from this data, yet existing methods largely overlook activity-level annotations, which capture temporal relationships and enrich semantic understanding. In this work, we introduce LSC-ADL, an ADL-annotated lifelog dataset derived from the LSC dataset, incorporating Activities of Daily Living (ADLs) as a structured semantic layer. Using a semi-automatic approach featuring the HDBSCAN algorithm for intra-class clustering and human-in-the-loop verification, we generate accurate ADL annotations to enhance retrieval explainability. By integrating action recognition into lifelog retrieval, LSC-ADL bridges a critical gap in existing research, offering a more context-aware representation of daily life. We believe this dataset will advance research in lifelog retrieval, activity recognition, and egocentric vision, ultimately improving the accuracy and interpretability of retrieved content. The ADL annotations can be downloaded at https://bit.ly/lsc-adl-annotations.
Egocentric video has seen increased interest in recent years, as it is used in a range of areas. However, most existing datasets are limited to a single perspective. In this paper, we present the CASTLE 2024 dataset, a multimodal collection containing ego- and exo-centric (i.e., first- and third-person perspective) video and audio from 15 time-aligned sources, as well as other sensor streams and auxiliary data. The dataset was recorded by volunteer participants over four days in a fixed location and includes the point of view of 10 participants, with an additional 5 fixed cameras providing an exocentric perspective. The entire dataset contains over 600 hours of UHD video recorded at 50 frames per second. In contrast to other datasets, CASTLE 2024 does not contain any partial censoring, such as blurred faces or distorted audio. The dataset is available via https://castle-dataset.github.io/.
This report presents the results of the 13th Video Browser Showdown, held at the 2024 International Conference on Multimedia Modeling on the 29th of January 2024 in Amsterdam, the Netherlands.
Personal data includes the digital footprints that we leave behind as part of our everyday activities, both online and offline in the real world. It includes data we collect ourselves, such as from wearables, as well as the data collected by others about our online behaviour and activities. Sometimes we are able to use the personal data we ourselves collect, in order to examine some parts of our lives but for the most part, our personal data is leveraged by third parties including internet companies, for services like targeted advertising and recommendations. Lifelogging is a form of extreme personal data gathering and in this article we present an overview of the tools used to manage access to lifelogs as demonstrated at the most recent of the annual Lifelog Search Challenge benchmarking workshops. Here, experimental systems are showcased in live, real time information seeking tasks by real users. This overview of these systems' capabilities show the range of possibilities for accessing our own personal data which may, in time, become more easily available as consumer-level services.
Recent years have witnessed an increasing amount of dialogue/conversation on the web especially on social media. That inspires the development of dialogue-based retrieval, in which retrieving videos based on dialogue is of increasing interest for recommendation systems. Different from other video retrieval tasks, dialogue-to-video retrieval uses structured queries in the form of user-generated dialogue as the search descriptor. We present a novel dialogue-to-video retrieval system, incorporating structured conversational information. Experiments conducted on the AVSD dataset show that our proposed approach using plain-text queries improves over the previous counterpart model by 15.8% on R@1. Furthermore, our approach using dialogue as a query, improves retrieval performance by 4.2%, 6.2%, 8.6% on R@1, R@5 and R@10 and outperforms the state-of-the-art model by 0.7%, 3.6% and 6.0% on R@1, R@5 and R@10 respectively.
Many models have been proposed for vision and language tasks, especially the image-text retrieval task. All state-of-the-art (SOTA) models in this challenge contained hundreds of millions of parameters. They also were pretrained on a large external dataset that has been proven to make a big improvement in overall performance. It is not easy to propose a new model with a novel architecture and intensively train it on a massive dataset with many GPUs to surpass many SOTA models, which are already available to use on the Internet. In this paper, we proposed a compact graph-based framework, named HADA, which can combine pretrained models to produce a better result, rather than building from scratch. First, we created a graph structure in which the nodes were the features extracted from the pretrained models and the edges connecting them. The graph structure was employed to capture and fuse the information from every pretrained model with each other. Then a graph neural network was applied to update the connection between the nodes to get the representative embedding vector for an image and text. Finally, we used the cosine similarity to match images with their relevant texts and vice versa to ensure a low inference time. Our experiments showed that, although HADA contained a tiny number of trainable parameters, it could increase baseline performance by more than 3.6% in terms of evaluation metrics in the Flickr30k dataset. Additionally, the proposed model did not train on any external dataset and did not require many GPUs but only 1 to train due to its small number of parameters. The source code is available at https://github.com/m2man/HADA.
Stress is a complex issue with wide-ranging physical and psychological impacts on human daily performance. Specifically, acute stress detection is becoming a valuable application in contextual human understanding. Two common approaches to training a stress detection model are subject-dependent and subject-independent training methods. Although subject-dependent training methods have proven to be the most accurate approach to build stress detection models, subject-independent models are a more practical and cost-efficient method, as they allow for the deployment of stress level detection and management systems in consumer-grade wearable devices without requiring training data for the end-user. To improve the performance of subject-independent stress detection models, in this paper, we introduce a stress-related bio-signal processing pipeline with a simple neural network architecture using statistical features extracted from multimodal contextual sensing sources including Electrodermal Activity (EDA), Blood Volume Pulse (BVP), and Skin Temperature (ST) captured from a consumer-grade wearable device. Using our proposed model architecture, we compare the accuracy between stress detection models that use measures from each individual signal source, and one model employing the fusion of multiple sensor sources. Extensive experiments on the publicly available WESAD dataset demonstrate that our proposed model outperforms conventional methods as well as providing 1.63% higher mean accuracy score compared to the state-of-the-art model while maintaining a low standard deviation. Our experiments also show that combining features from multiple sources produce more accurate predictions than using only one sensor source individually.