Dima Damen

Kristen Grauman, Andrew Westbury, Lorenzo Torresani et al. · cmu, gatech

We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured egocentric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions -- including a novel "expert commentary" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled human activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community. Project page: http://ego-exo4d-data.org/

Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan et al. · microsoft-research, uw

Video-Language Pretraining (VLP), which aims to learn transferable representation to advance a wide range of video-text downstream tasks, has recently received increasing attention. Best performing works rely on large-scale, 3rd-person video-text datasets, such as HowTo100M. In this work, we exploit the recently released Ego4D dataset to pioneer Egocentric VLP along three directions. (i) We create EgoClip, a 1st-person video-text pretraining dataset comprising 3.8M clip-text pairs well-chosen from Ego4D, covering a large variety of human daily activities. (ii) We propose a novel pretraining objective, dubbed EgoNCE, which adapts video-text contrastive learning to the egocentric domain by mining egocentric-aware positive and negative samples. (iii) We introduce EgoMCQ, a development benchmark that is close to EgoClip and hence can support effective validation and fast exploration of our design decisions in EgoClip and EgoNCE. Furthermore, we demonstrate strong performance on five egocentric downstream tasks across three datasets: video-text retrieval on EPIC-KITCHENS-100; action recognition on Charades-Ego; natural language query, moment query, and object state change classification on Ego4D challenge benchmarks. The dataset and code are available at https://github.com/showlab/EgoVLP.

Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan et al. · microsoft-research, uw

In this report, we propose a video-language pretraining (VLP) based solution \cite{kevin2022egovlp} for four Ego4D challenge tasks, including Natural Language Query (NLQ), Moment Query (MQ), Object State Change Classification (OSCC), and PNR Localization (PNR). Especially, we exploit the recently released Ego4D dataset \cite{grauman2021ego4d} to pioneer Egocentric VLP from pretraining dataset, pretraining objective, and development set. Based on the above three designs, we develop a pretrained video-language model that is able to transfer its egocentric video-text representation or video-only representation to several video downstream tasks. Our Egocentric VLP achieves 10.46R@1&IoU @0.3 on NLQ, 10.33 mAP on MQ, 74% Acc on OSCC, 0.67 sec error on PNR. The code is available at https://github.com/showlab/EgoVLP.

Ahmad Darkhalil, Dandan Shan, Bin Zhu et al.

We introduce VISOR, a new dataset of pixel annotations and a benchmark suite for segmenting hands and active objects in egocentric video. VISOR annotates videos from EPIC-KITCHENS, which comes with a new set of challenges not encountered in current video segmentation datasets. Specifically, we need to ensure both short- and long-term consistency of pixel-level annotations as objects undergo transformative interactions, e.g. an onion is peeled, diced and cooked - where we aim to obtain accurate pixel-level annotations of the peel, onion pieces, chopping board, knife, pan, as well as the acting hands. VISOR introduces an annotation pipeline, AI-powered in parts, for scalability and quality. In total, we publicly release 272K manual semantic masks of 257 object classes, 9.9M interpolated dense masks, 67K hand-object relations, covering 36 hours of 179 untrimmed videos. Along with the annotations, we introduce three challenges in video object segmentation, interaction understanding and long-term reasoning. For data, code and leaderboards: http://epic-kitchens.github.io/VISOR

Dena Bazazian, Andrew Calway, Dima Damen

We introduce a segmentation-guided approach to synthesise images that integrate features from two distinct domains. Images synthesised by our dual-domain model belong to one domain within the semantic mask, and to another in the rest of the image - smoothly integrated. We build on the successes of few-shot StyleGAN and single-shot semantic segmentation to minimise the amount of training required in utilising two domains. The method combines a few-shot cross-domain StyleGAN with a latent optimiser to achieve images containing features of two distinct domains. We use a segmentation-guided perceptual loss, which compares both pixel-level and activations between domain-specific and dual-domain synthetic images. Results demonstrate qualitatively and quantitatively that our model is capable of synthesising dual-domain images on a variety of objects (faces, horses, cats, cars), domains (natural, caricature, sketches) and part-based masks (eyes, nose, mouth, hair, car bonnet). The code is publicly available at: https://github.com/denabazazian/Dual-Domain-Synthesis.

Jaesung Huh, Jacob Chalk, Evangelos Kazakos et al.

We introduce EPIC-SOUNDS, a large-scale dataset of audio annotations capturing temporal extents and class labels within the audio stream of the egocentric videos. We propose an annotation pipeline where annotators temporally label distinguishable audio segments and describe the action that could have caused this sound. We identify actions that can be discriminated purely from audio, through grouping these free-form descriptions of audio into classes. For actions that involve objects colliding, we collect human annotations of the materials of these objects (e.g. a glass object being placed on a wooden surface), which we verify from video, discarding ambiguities. Overall, EPIC-SOUNDS includes 78.4k categorised segments of audible events and actions, distributed across 44 classes as well as 39.2k non-categorised segments. We train and evaluate state-of-the-art audio recognition and detection models on our dataset, for both audio-only and audio-visual methods. We also conduct analysis on: the temporal overlap between audio events, the temporal and label correlations between audio and visual modalities, the ambiguities in annotating materials from audio-only input, the importance of audio-only labels and the limitations of current models to understand actions that sound.

Hanyuan Wang, Majid Mirmehdi, Dima Damen et al.

Previous one-stage action detection approaches have modelled temporal dependencies using only the visual modality. In this paper, we explore different strategies to incorporate the audio modality, using multi-scale cross-attention to fuse the two modalities. We also demonstrate the correlation between the distance from the timestep to the action centre and the accuracy of the predicted boundaries. Thus, we propose a novel network head to estimate the closeness of timesteps to the action centre, which we call the centricity score. This leads to increased confidence for proposals that exhibit more precise boundaries. Our method can be integrated with other one-stage anchor-free architectures and we demonstrate this on three recent baselines on the EPIC-Kitchens-100 action detection benchmark where we achieve state-of-the-art performance. Detailed ablation studies showcase the benefits of fusing audio and our proposed centricity scores. Code and models for our proposed method are publicly available at https://github.com/hanielwang/Audio-Visual-TAD.git

Kevin Flanagan, Dima Damen, Michael Wray

The onset of long-form egocentric datasets such as Ego4D and EPIC-Kitchens presents a new challenge for the task of Temporal Sentence Grounding (TSG). Compared to traditional benchmarks on which this task is evaluated, these datasets offer finer-grained sentences to ground in notably longer videos. In this paper, we develop an approach for learning to ground sentences in these datasets using only narrations and their corresponding rough narration timestamps. We propose to artificially merge clips to train for temporal grounding in a contrastive manner using text-conditioning attention. This Clip Merging (CliMer) approach is shown to be effective when compared with a high performing TSG method -- e.g. mean R@1 improves from 3.9 to 5.7 on Ego4D and from 10.7 to 13.0 on EPIC-Kitchens. Code and data splits available from: https://github.com/keflanagan/CliMer

Vadim Tschernezki, Ahmad Darkhalil, Zhifan Zhu et al.

Neural rendering is fuelling a unification of learning, 3D geometry and video understanding that has been waiting for more than two decades. Progress, however, is still hampered by a lack of suitable datasets and benchmarks. To address this gap, we introduce EPIC Fields, an augmentation of EPIC-KITCHENS with 3D camera information. Like other datasets for neural rendering, EPIC Fields removes the complex and expensive step of reconstructing cameras using photogrammetry, and allows researchers to focus on modelling problems. We illustrate the challenge of photogrammetry in egocentric videos of dynamic actions and propose innovations to address them. Compared to other neural rendering datasets, EPIC Fields is better tailored to video understanding because it is paired with labelled action segments and the recent VISOR segment annotations. To further motivate the community, we also evaluate two benchmark tasks in neural rendering and segmenting dynamic objects, with strong baselines that showcase what is not possible today. We also highlight the advantage of geometry in semi-supervised video object segmentations on the VISOR annotations. EPIC Fields reconstructs 96% of videos in EPICKITCHENS, registering 19M frames in 99 hours recorded in 45 kitchens.

Chiara Plizzari, Gabriele Goletto, Antonino Furnari et al.

What will the future be? We wonder! In this survey, we explore the gap between current research in egocentric vision and the ever-anticipated future, where wearable computing, with outward facing cameras and digital overlays, is expected to be integrated in our every day lives. To understand this gap, the article starts by envisaging the future through character-based stories, showcasing through examples the limitations of current technology. We then provide a mapping between this future and previously defined research tasks. For each task, we survey its seminal works, current state-of-the-art methodologies and available datasets, then reflect on shortcomings that limit its applicability to future research. Note that this survey focuses on software models for egocentric vision, independent of any specific hardware. The paper concludes with recommendations for areas of immediate explorations so as to unlock our path to the future always-on, personalised and life-enhancing egocentric vision.

Chiara Plizzari, Toby Perrett, Barbara Caputo et al.

We propose and address a new generalisation problem: can a model trained for action recognition successfully classify actions when they are performed within a previously unseen scenario and in a previously unseen location? To answer this question, we introduce the Action Recognition Generalisation Over scenarios and locations dataset (ARGO1M), which contains 1.1M video clips from the large-scale Ego4D dataset, across 10 scenarios and 13 locations. We demonstrate recognition models struggle to generalise over 10 proposed test splits, each of an unseen scenario in an unseen location. We thus propose CIR, a method to represent each video as a Cross-Instance Reconstruction of videos from other domains. Reconstructions are paired with text narrations to guide the learning of a domain generalisable representation. We provide extensive analysis and ablations on ARGO1M that show CIR outperforms prior domain generalisation works on all test splits. Code and data: https://chiaraplizz.github.io/what-can-a-cook/.

Toby Perrett, Saptarshi Sinha, Tilo Burghardt et al.

This paper presents an investigation into long-tail video recognition. We demonstrate that, unlike naturally-collected video datasets and existing long-tail image benchmarks, current video benchmarks fall short on multiple long-tailed properties. Most critically, they lack few-shot classes in their tails. In response, we propose new video benchmarks that better assess long-tail recognition, by sampling subsets from two datasets: SSv2 and VideoLT. We then propose a method, Long-Tail Mixed Reconstruction, which reduces overfitting to instances from few-shot classes by reconstructing them as weighted combinations of samples from head classes. LMR then employs label mixing to learn robust decision boundaries. It achieves state-of-the-art average class accuracy on EPIC-KITCHENS and the proposed SSv2-LT and VideoLT-LT. Benchmarks and code at: tobyperrett.github.io/lmr

Gabriele Goletto, Tushar Nagarajan, Giuseppe Averta et al.

Egocentric videos provide a unique perspective into individuals' daily experiences, yet their unstructured nature presents challenges for perception. In this paper, we introduce AMEGO, a novel approach aimed at enhancing the comprehension of very-long egocentric videos. Inspired by the human's ability to maintain information from a single watching, AMEGO focuses on constructing a self-contained representations from one egocentric video, capturing key locations and object interactions. This representation is semantic-free and facilitates multiple queries without the need to reprocess the entire visual content. Additionally, to evaluate our understanding of very-long egocentric videos, we introduce the new Active Memories Benchmark (AMB), composed of more than 20K of highly challenging visual queries from EPIC-KITCHENS. These queries cover different levels of video reasoning (sequencing, concurrency and temporal grounding) to assess detailed video understanding capabilities. We showcase improved performance of AMEGO on AMB, surpassing other video QA baselines by a substantial margin.

Alexandros Stergiou, Dima Damen

Early action prediction deals with inferring the ongoing action from partially-observed videos, typically at the outset of the video. We propose a bottleneck-based attention model that captures the evolution of the action, through progressive sampling over fine-to-coarse scales. Our proposed Temporal Progressive (TemPr) model is composed of multiple attention towers, one for each scale. The predicted action label is based on the collective agreement considering confidences of these towers. Extensive experiments over four video datasets showcase state-of-the-art performance on the task of Early Action Prediction across a range of encoder architectures. We demonstrate the effectiveness and consistency of TemPr through detailed ablations.

Alexandros Stergiou, Dima Damen

A key function of auditory cognition is the association of characteristic sounds with their corresponding semantics over time. Humans attempting to discriminate between fine-grained audio categories, often replay the same discriminative sounds to increase their prediction confidence. We propose an end-to-end attention-based architecture that through selective repetition attends over the most discriminative sounds across the audio sequence. Our model initially uses the full audio sequence and iteratively refines the temporal segments replayed based on slot attention. At each playback, the selected segments are replayed using a smaller hop length which represents higher resolution features within these segments. We show that our method can consistently achieve state-of-the-art performance across three audio-classification benchmarks: AudioSet, VGG-Sound, and EPIC-KITCHENS-100.

Adriano Fragomeni, Michael Wray, Dima Damen

In this paper, we re-examine the task of cross-modal clip-sentence retrieval, where the clip is part of a longer untrimmed video. When the clip is short or visually ambiguous, knowledge of its local temporal context (i.e. surrounding video segments) can be used to improve the retrieval performance. We propose Context Transformer (ConTra); an encoder architecture that models the interaction between a video clip and its local temporal context in order to enhance its embedded representations. Importantly, we supervise the context transformer using contrastive losses in the cross-modal embedding space. We explore context transformers for video and text modalities. Results consistently demonstrate improved performance on three datasets: YouCook2, EPIC-KITCHENS and a clip-sentence version of ActivityNet Captions. Exhaustive ablation studies and context analysis show the efficacy of the proposed method.

Hanyuan Wang, Majid Mirmehdi, Dima Damen et al.

Current one-stage action detection methods, which simultaneously predict action boundaries and the corresponding class, do not estimate or use a measure of confidence in their boundary predictions, which can lead to inaccurate boundaries. We incorporate the estimation of boundary confidence into one-stage anchor-free detection, through an additional prediction head that predicts the refined boundaries with higher confidence. We obtain state-of-the-art performance on the challenging EPIC-KITCHENS-100 action detection as well as the standard THUMOS14 action detection benchmarks, and achieve improvement on the ActivityNet-1.3 benchmark.

Masashi Hatano, Saptarshi Sinha, Jacob Chalk et al.

Human motion generation is a challenging task that aims to create realistic motion imitating natural human behaviour. We focus on the well-studied behaviour of priming an object/location for pick up or put down -- that is, the spotting of an object/location from a distance, known as gaze priming, followed by the motion of approaching and reaching the target location. To that end, we curate, for the first time, 23.7K gaze-primed human motion sequences for reaching target object locations from five publicly available datasets, i.e., HD-EPIC, MoGaze, HOT3D, ADT, and GIMO. We pre-train a text-conditioned diffusion-based motion generation model, then fine-tune it conditioned on goal pose or location, on our curated sequences. Importantly, we evaluate the ability of the generated motion to imitate natural human movement through several metrics, including the 'Reach Success' and a newly introduced 'Prime Success' metric. On the largest dataset, HD-EPIC, our model achieves 60% prime success and 89% reach success when conditioned on the goal object location.

Alessandro Masullo, Toby Perrett, Tilo Burghardt et al.

We propose a novel approach to multimodal sensor fusion for Ambient Assisted Living (AAL) which takes advantage of learning using privileged information (LUPI). We address two major shortcomings of standard multimodal approaches, limited area coverage and reduced reliability. Our new framework fuses the concept of modality hallucination with triplet learning to train a model with different modalities to handle missing sensors at inference time. We evaluate the proposed model on inertial data from a wearable accelerometer device, using RGB videos and skeletons as privileged modalities, and show an improvement of accuracy of an average 6.6% on the UTD-MHAD dataset and an average 5.5% on the Berkeley MHAD dataset, reaching a new state-of-the-art for inertial-only classification accuracy on these datasets. We validate our framework through several ablation studies.

Tengda Han, Sayna Ebrahimi, Dilara Gokay et al.

We introduce the "single-life" learning paradigm, where we train a distinct vision model exclusively on egocentric videos captured by one individual. We leverage the multiple viewpoints naturally captured within a single life to learn a visual encoder in a self-supervised manner. Our experiments demonstrate three key findings. First, models trained independently on different lives develop a highly aligned geometric understanding. We demonstrate this by training visual encoders on distinct datasets each capturing a different life, both indoors and outdoors, as well as introducing a novel cross-attention-based metric to quantify the functional alignment of the internal representations developed by different models. Second, we show that single-life models learn generalizable geometric representations that effectively transfer to downstream tasks, such as depth estimation, in unseen environments. Third, we demonstrate that training on up to 30 hours from one week of the same person's life leads to comparable performance to training on 30 hours of diverse web data, highlighting the strength of single-life representation learning. Overall, our results establish that the shared structure of the world, both leads to consistency in models trained on individual lives, and provides a powerful signal for visual representation learning.

Zhifan Zhu, Yifei Huang, Yoichi Sato et al.

Humans can intuitively parallelise complex activities, but can a model learn this from observing a single person? Given one egocentric video, we introduce the N-Body Problem: how N individuals, can hypothetically perform the same set of tasks observed in this video. The goal is to maximise speed-up, but naive assignment of video segments to individuals often violates real-world constraints, leading to physically impossible scenarios like two people using the same object or occupying the same space. To address this, we formalise the N-Body Problem and propose a suite of metrics to evaluate both performance (speed-up, task coverage) and feasibility (spatial collisions, object conflicts and causal constraints). We then introduce a structured prompting strategy that guides a Vision-Language Model (VLM) to reason about the 3D environment, object usage, and temporal dependencies to produce a viable parallel execution. On 100 videos from EPIC-Kitchens and HD-EPIC, our method for N = 2 boosts action coverage by 45% over a baseline prompt for Gemini 2.5 Pro, while simultaneously slashing collision rates, object and causal conflicts by 55%, 45% and 55% respectively.

David Pujol-Perich, Albert Clapés, Dima Damen et al.

In this work, we investigate the degradation of existing VMR methods, particularly of DETR architectures, when trained on caption-based queries but evaluated on search queries. For this, we introduce three benchmarks by modifying the textual queries in three public VMR datasets -- i.e., HD-EPIC, YouCook2 and ActivityNet-Captions. Our analysis reveals two key generalization challenges: (i) A language gap, arising from the linguistic under-specification of search queries, and (ii) a multi-moment gap, caused by the shift from single-moment to multi-moment queries. We also identify a critical issue in these architectures -- an active decoder-query collapse -- as a primary cause of the poor generalization to multi-moment instances. We mitigate this issue with architectural modifications that effectively increase the number of active decoder queries. Extensive experiments demonstrate that our approach improves performance on search queries by up to 14.82% mAP_m, and up to 21.83% mAP_m on multi-moment search queries. The code, models and data are available in the project webpage: https://davidpujol.github.io/beyond-vmr/

Joseph Heyward, Nikhil Pathasarathy, Tyler Zhu et al.

The Third Perception Test challenge was organised as a full-day workshop alongside the IEEE/CVF International Conference on Computer Vision (ICCV) 2025. Its primary goal is to benchmark state-of-the-art video models and measure the progress in multimodal perception. This year, the workshop featured 2 guest tracks as well: KiVA (an image understanding challenge) and Physic-IQ (a video generation challenge). In this report, we summarise the results from the main Perception Test challenge, detailing both the existing tasks as well as novel additions to the benchmark. In this iteration, we placed an emphasis on task unification, as this poses a more challenging test for current SOTA multimodal models. The challenge included five consolidated tracks: unified video QA, unified object and point tracking, unified action and sound localisation, grounded video QA, and hour-long video QA, alongside an analysis and interpretability track that is still open for submissions. Notably, the unified video QA track introduced a novel subset that reformulates traditional perception tasks (such as point tracking and temporal action localisation) as multiple-choice video QA questions that video-language models can natively tackle. The unified object and point tracking merged the original object tracking and point tracking tasks, whereas the unified action and sound localisation merged the original temporal action localisation and temporal sound localisation tracks. Accordingly, we required competitors to use unified approaches rather than engineered pipelines with task-specific models. By proposing such a unified challenge, Perception Test 2025 highlights the significant difficulties existing models face when tackling diverse perception tasks through unified interfaces.

Zihui Xue, Kristen Grauman, Dima Damen et al.

Can one perceive a video's content without seeing its pixels, just from the camera trajectory-the path it carves through space? This paper is the first to systematically investigate this seemingly implausible question. Towards this end, we propose a contrastive learning framework to train CamFormer, a dedicated encoder that projects camera pose trajectories into a joint embedding space, aligning them with natural language. We find that, contrary to its apparent simplicity, the camera trajectory is a remarkably informative signal to uncover video content. In other words, "how you move" can indeed reveal "what you are doing" (egocentric) or "observing" (exocentric). We demonstrate the versatility of our learned CamFormer embeddings on a diverse suite of downstream tasks, ranging from cross-modal alignment to classification and temporal analysis. Importantly, our representations are robust across diverse camera pose estimation methods, including both high-fidelity multi-sensored and standard RGB-only estimators. Our findings establish camera trajectory as a lightweight, robust, and versatile modality for perceiving video content.

Zhifan Zhu, Siddhant Bansal, Shashank Tripathi et al.

We introduce the task of Reconstructing Objects along Hand Interaction Timelines (ROHIT). We first define the Hand Interaction Timeline (HIT) from a rigid object's perspective. In a HIT, an object is first static relative to the scene, then is held in hand following contact, where its pose changes. This is usually followed by a firm grip during use, before it is released to be static again w.r.t. to the scene. We model these pose constraints over the HIT, and propose to propagate the object's pose along the HIT enabling superior reconstruction using our proposed Constrained Optimisation and Propagation (COP) framework. Importantly, we focus on timelines with stable grasps - i.e. where the hand is stably holding an object, effectively maintaining constant contact during use. This allows us to efficiently annotate, study, and evaluate object reconstruction in videos without 3D ground truth. We evaluate our proposed task, ROHIT, over two egocentric datasets, HOT3D and in-the-wild EPIC-Kitchens. In HOT3D, we curate 1.2K clips of stable grasps. In EPIC-Kitchens, we annotate 2.4K clips of stable grasps including 390 object instances across 9 categories from videos of daily interactions in 141 environments. Without 3D ground truth, we utilise 2D projection error to assess the reconstruction. Quantitatively, COP improves stable grasp reconstruction by 6.2-11.3% and HIT reconstruction by up to 24.5% with constrained pose propagation.

Rhodri Guerrier, Adam W. Harley, Dima Damen

Recent advances in foundational 3D reconstruction models, such as DUSt3R and MASt3R, have shown great potential in 2D and 3D correspondence in static scenes. In this paper, we propose to adapt them for the task of point tracking through 3D grounded correspondence. We first demonstrate that these models are competitive point trackers when focusing on static points, present in current point tracking benchmarks ($+33.5\%$ on EgoPoints vs. CoTracker2). We propose to combine the reconstruction loss with training for dynamic correspondence along with a visibility head, and fine-tuning MASt3R for point tracking using a relatively small amount of synthetic data. Importantly, we only train and evaluate on pairs of frames where one contains the query point, effectively removing any temporal context. Using a mix of dynamic and static point correspondences, we achieve competitive or superior point tracking results on four datasets (e.g. competitive on TAP-Vid-DAVIS 73.8 $δ_{avg}$ / 85.8\% occlusion acc. for PointSt3R compared to 75.7 / 88.3\% for CoTracker2; and significantly outperform CoTracker3 on EgoPoints 61.3 vs 54.2 and RGB-S 87.0 vs 82.8). We also present results on 3D point tracking along with several ablations on training datasets and percentage of dynamic correspondences.

Valentin Popescu, Dima Damen, Toby Perrett

In this paper, we evaluate state-of-the-art OCR methods on Egocentric data. We annotate text in EPIC-KITCHENS images, and demonstrate that existing OCR methods struggle with rotated text, which is frequently observed on objects being handled. We introduce a simple rotate-and-merge procedure which can be applied to pre-trained OCR models that halves the normalized edit distance error. This suggests that future OCR attempts should incorporate rotation into model design and training procedures.

Jacob Chalk, Jaesung Huh, Evangelos Kazakos et al.

Diverse actions give rise to rich audio-visual signals in long videos. Recent works showcase that the two modalities of audio and video exhibit different temporal extents of events and distinct labels. We address the interplay between the two modalities in long videos by explicitly modelling the temporal extents of audio and visual events. We propose the Time Interval Machine (TIM) where a modality-specific time interval poses as a query to a transformer encoder that ingests a long video input. The encoder then attends to the specified interval, as well as the surrounding context in both modalities, in order to recognise the ongoing action. We test TIM on three long audio-visual video datasets: EPIC-KITCHENS, Perception Test, and AVE, reporting state-of-the-art (SOTA) for recognition. On EPIC-KITCHENS, we beat previous SOTA that utilises LLMs and significantly larger pre-training by 2.9% top-1 action recognition accuracy. Additionally, we show that TIM can be adapted for action detection, using dense multi-scale interval queries, outperforming SOTA on EPIC-KITCHENS-100 for most metrics, and showing strong performance on the Perception Test. Our ablations show the critical role of integrating the two modalities and modelling their time intervals in achieving this performance. Code and models at: https://github.com/JacobChalk/TIM

João Carreira, Dilara Gokay, Michael King et al.

Scaling has not yet been convincingly demonstrated for pure self-supervised learning from video. However, prior work has focused evaluations on semantic-related tasks $\unicode{x2013}$ action classification, ImageNet classification, etc. In this paper we focus on evaluating self-supervised learning on non-semantic vision tasks that are more spatial (3D) and temporal (+1D = 4D), such as camera pose estimation, point and object tracking, and depth estimation. We show that by learning from very large video datasets, masked auto-encoding (MAE) with transformer video models actually scales, consistently improving performance on these 4D tasks, as model size increases from 20M all the way to the largest by far reported self-supervised video model $\unicode{x2013}$ 22B parameters. Rigorous apples-to-apples comparison with many recent image and video models demonstrates the benefits of scaling 4D representations. Pretrained models are available at https://github.com/google-deepmind/representations4d .

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Tomáš Souček, Prajwal Gatti, Michael Wray et al.

The goal of this work is to generate step-by-step visual instructions in the form of a sequence of images, given an input image that provides the scene context and the sequence of textual instructions. This is a challenging problem as it requires generating multi-step image sequences to achieve a complex goal while being grounded in a specific environment. Part of the challenge stems from the lack of large-scale training data for this problem. The contribution of this work is thus three-fold. First, we introduce an automatic approach for collecting large step-by-step visual instruction training data from instructional videos. We apply this approach to one million videos and create a large-scale, high-quality dataset of 0.6M sequences of image-text pairs. Second, we develop and train ShowHowTo, a video diffusion model capable of generating step-by-step visual instructions consistent with the provided input image. Third, we evaluate the generated image sequences across three dimensions of accuracy (step, scene, and task) and show our model achieves state-of-the-art results on all of them. Our code, dataset, and trained models are publicly available.

Kevin Flanagan, Dima Damen, Michael Wray

Video Moment Retrieval is a common task to evaluate the performance of visual-language models - it involves localising start and end times of moments in videos from query sentences. The current task formulation assumes that the queried moment is present in the video, resulting in false positive moment predictions when irrelevant query sentences are provided. In this paper we propose the task of Negative-Aware Video Moment Retrieval (NA-VMR), which considers both moment retrieval accuracy and negative query rejection accuracy. We make the distinction between In-Domain and Out-of-Domain negative queries and provide new evaluation benchmarks for two popular video moment retrieval datasets: QVHighlights and Charades-STA. We analyse the ability of current SOTA video moment retrieval approaches to adapt to Negative-Aware Video Moment Retrieval and propose UniVTG-NA, an adaptation of UniVTG designed to tackle NA-VMR. UniVTG-NA achieves high negative rejection accuracy (avg. $98.4\%$) scores while retaining moment retrieval scores to within $3.87\%$ Recall@1. Dataset splits and code are available at https://github.com/keflanagan/MomentofUntruth

Viorica Pătrăucean, Lucas Smaira, Ankush Gupta et al.

We propose a novel multimodal video benchmark - the Perception Test - to evaluate the perception and reasoning skills of pre-trained multimodal models (e.g. Flamingo, SeViLA, or GPT-4). Compared to existing benchmarks that focus on computational tasks (e.g. classification, detection or tracking), the Perception Test focuses on skills (Memory, Abstraction, Physics, Semantics) and types of reasoning (descriptive, explanatory, predictive, counterfactual) across video, audio, and text modalities, to provide a comprehensive and efficient evaluation tool. The benchmark probes pre-trained models for their transfer capabilities, in a zero-shot / few-shot or limited finetuning regime. For these purposes, the Perception Test introduces 11.6k real-world videos, 23s average length, designed to show perceptually interesting situations, filmed by around 100 participants worldwide. The videos are densely annotated with six types of labels (multiple-choice and grounded video question-answers, object and point tracks, temporal action and sound segments), enabling both language and non-language evaluations. The fine-tuning and validation splits of the benchmark are publicly available (CC-BY license), in addition to a challenge server with a held-out test split. Human baseline results compared to state-of-the-art video QA models show a substantial gap in performance (91.4% vs 46.2%), suggesting that there is significant room for improvement in multimodal video understanding. Dataset, baseline code, and challenge server are available at https://github.com/deepmind/perception_test

Hanyuan Wang, Dima Damen, Majid Mirmehdi et al.

We propose a Temporal Voting Network (TVNet) for action localization in untrimmed videos. This incorporates a novel Voting Evidence Module to locate temporal boundaries, more accurately, where temporal contextual evidence is accumulated to predict frame-level probabilities of start and end action boundaries. Our action-independent evidence module is incorporated within a pipeline to calculate confidence scores and action classes. We achieve an average mAP of 34.6% on ActivityNet-1.3, particularly outperforming previous methods with the highest IoU of 0.95. TVNet also achieves mAP of 56.0% when combined with PGCN and 59.1% with MUSES at 0.5 IoU on THUMOS14 and outperforms prior work at all thresholds. Our code is available at https://github.com/hanielwang/TVNet.

Evangelos Kazakos, Jaesung Huh, Arsha Nagrani et al.

In egocentric videos, actions occur in quick succession. We capitalise on the action's temporal context and propose a method that learns to attend to surrounding actions in order to improve recognition performance. To incorporate the temporal context, we propose a transformer-based multimodal model that ingests video and audio as input modalities, with an explicit language model providing action sequence context to enhance the predictions. We test our approach on EPIC-KITCHENS and EGTEA datasets reporting state-of-the-art performance. Our ablations showcase the advantage of utilising temporal context as well as incorporating audio input modality and language model to rescore predictions. Code and models at: https://github.com/ekazakos/MTCN.

Will Price, Dima Damen

We benchmark contemporary action recognition models (TSN, TRN, and TSM) on the recently introduced EPIC-Kitchens dataset and release pretrained models on GitHub (https://github.com/epic-kitchens/action-models) for others to build upon. In contrast to popular action recognition datasets like Kinetics, Something-Something, UCF101, and HMDB51, EPIC-Kitchens is shot from an egocentric perspective and captures daily actions in-situ. In this report, we aim to understand how well these models can tackle the challenges present in this dataset, such as its long tail class distribution, unseen environment test set, and multiple tasks (verb, noun and, action classification). We discuss the models' shortcomings and avenues for future research.

Tomáš Souček, Dima Damen, Michael Wray et al.

We address the task of generating temporally consistent and physically plausible images of actions and object state transformations. Given an input image and a text prompt describing the targeted transformation, our generated images preserve the environment and transform objects in the initial image. Our contributions are threefold. First, we leverage a large body of instructional videos and automatically mine a dataset of triplets of consecutive frames corresponding to initial object states, actions, and resulting object transformations. Second, equipped with this data, we develop and train a conditioned diffusion model dubbed GenHowTo. Third, we evaluate GenHowTo on a variety of objects and actions and show superior performance compared to existing methods. In particular, we introduce a quantitative evaluation where GenHowTo achieves 88% and 74% on seen and unseen interaction categories, respectively, outperforming prior work by a large margin.

Toby Perrett, Ahmad Darkhalil, Saptarshi Sinha et al.

We present a validation dataset of newly-collected kitchen-based egocentric videos, manually annotated with highly detailed and interconnected ground-truth labels covering: recipe steps, fine-grained actions, ingredients with nutritional values, moving objects, and audio annotations. Importantly, all annotations are grounded in 3D through digital twinning of the scene, fixtures, object locations, and primed with gaze. Footage is collected from unscripted recordings in diverse home environments, making HDEPIC the first dataset collected in-the-wild but with detailed annotations matching those in controlled lab environments. We show the potential of our highly-detailed annotations through a challenging VQA benchmark of 26K questions assessing the capability to recognise recipes, ingredients, nutrition, fine-grained actions, 3D perception, object motion, and gaze direction. The powerful long-context Gemini Pro only achieves 38.5% on this benchmark, showcasing its difficulty and highlighting shortcomings in current VLMs. We additionally assess action recognition, sound recognition, and long-term video-object segmentation on HD-EPIC. HD-EPIC is 41 hours of video in 9 kitchens with digital twins of 413 kitchen fixtures, capturing 69 recipes, 59K fine-grained actions, 51K audio events, 20K object movements and 37K object masks lifted to 3D. On average, we have 263 annotations per minute of our unscripted videos.

Chiara Plizzari, Shubham Goel, Toby Perrett et al. · berkeley

As humans move around, performing their daily tasks, they are able to recall where they have positioned objects in their environment, even if these objects are currently out of their sight. In this paper, we aim to mimic this spatial cognition ability. We thus formulate the task of Out of Sight, Not Out of Mind - 3D tracking active objects using observations captured through an egocentric camera. We introduce a simple but effective approach to address this challenging problem, called Lift, Match, and Keep (LMK). LMK lifts partial 2D observations to 3D world coordinates, matches them over time using visual appearance, 3D location and interactions to form object tracks, and keeps these object tracks even when they go out-of-view of the camera. We benchmark LMK on 100 long videos from EPIC-KITCHENS. Our results demonstrate that spatial cognition is critical for correctly locating objects over short and long time scales. E.g., for one long egocentric video, we estimate the 3D location of 50 active objects. After 120 seconds, 57% of the objects are correctly localised by LMK, compared to just 33% by a recent 3D method for egocentric videos and 17% by a general 2D tracking method.

Saptarshi Sinha, Alexandros Stergiou, Dima Damen

Video repetition counting infers the number of repetitions of recurring actions or motion within a video. We propose an exemplar-based approach that discovers visual correspondence of video exemplars across repetitions within target videos. Our proposed Every Shot Counts (ESCounts) model is an attention-based encoder-decoder that encodes videos of varying lengths alongside exemplars from the same and different videos. In training, ESCounts regresses locations of high correspondence to the exemplars within the video. In tandem, our method learns a latent that encodes representations of general repetitive motions, which we use for exemplar-free, zero-shot inference. Extensive experiments over commonly used datasets (RepCount, Countix, and UCFRep) showcase ESCounts obtaining state-of-the-art performance across all three datasets. Detailed ablations further demonstrate the effectiveness of our method.

Zhifan Zhu, Dima Damen

We propose the task of Hand-Object Stable Grasp Reconstruction (HO-SGR), the reconstruction of frames during which the hand is stably holding the object. We first develop the stable grasp definition based on the intuition that the in-contact area between the hand and object should remain stable. By analysing the 3D ARCTIC dataset, we identify stable grasp durations and showcase that objects in stable grasps move within a single degree of freedom (1-DoF). We thereby propose a method to jointly optimise all frames within a stable grasp, minimising object motions to a latent 1-DoF. Finally, we extend the knowledge to in-the-wild videos by labelling 2.4K clips of stable grasps. Our proposed EPIC-Grasps dataset includes 390 object instances of 9 categories, featuring stable grasps from videos of daily interactions in 141 environments. Without 3D ground truth, we use stable contact areas and 2D projection masks to assess the HO-SGR task in the wild. We evaluate relevant methods and our approach preserves significantly higher stable contact area, on both EPIC-Grasps and stable grasp sub-sequences from the ARCTIC dataset.

Siddhant Bansal, Michael Wray, Dima Damen

Large Vision Language Models (VLMs) are now the de facto state-of-the-art for a number of tasks including visual question answering, recognising objects, and spatial referral. In this work, we propose the HOI-Ref task for egocentric images that aims to understand interactions between hands and objects using VLMs. To enable HOI-Ref, we curate the HOI-QA dataset that consists of 3.9M question-answer pairs for training and evaluating VLMs. HOI-QA includes questions relating to locating hands, objects, and critically their interactions (e.g. referring to the object being manipulated by the hand). We train the first VLM for HOI-Ref on this dataset and call it VLM4HOI. Our results demonstrate that VLMs trained for referral on third person images fail to recognise and refer hands and objects in egocentric images. When fine-tuned on our egocentric HOI-QA dataset, performance improves by 27.9% for referring hands and objects, and by 26.7% for referring interactions.

Ahmad Darkhalil, Rhodri Guerrier, Adam W. Harley et al.

We introduce EgoPoints, a benchmark for point tracking in egocentric videos. We annotate 4.7K challenging tracks in egocentric sequences. Compared to the popular TAP-Vid-DAVIS evaluation benchmark, we include 9x more points that go out-of-view and 59x more points that require re-identification (ReID) after returning to view. To measure the performance of models on these challenging points, we introduce evaluation metrics that specifically monitor tracking performance on points in-view, out-of-view, and points that require re-identification. We then propose a pipeline to create semi-real sequences, with automatic ground truth. We generate 11K such sequences by combining dynamic Kubric objects with scene points from EPIC Fields. When fine-tuning point tracking methods on these sequences and evaluating on our annotated EgoPoints sequences, we improve CoTracker across all metrics, including the tracking accuracy $δ^\star_{\text{avg}}$ by 2.7 percentage points and accuracy on ReID sequences (ReID$δ_{\text{avg}}$) by 2.4 points. We also improve $δ^\star_{\text{avg}}$ and ReID$δ_{\text{avg}}$ of PIPs++ by 0.3 and 2.8 respectively.

Tengda Han, Dilara Gokay, Joseph Heyward et al.

We address the challenge of representation learning from a continuous stream of video as input, in a self-supervised manner. This differs from the standard approaches to video learning where videos are chopped and shuffled during training in order to create a non-redundant batch that satisfies the independently and identically distributed (IID) sample assumption expected by conventional training paradigms. When videos are only available as a continuous stream of input, the IID assumption is evidently broken, leading to poor performance. We demonstrate the drop in performance when moving from shuffled to sequential learning on three tasks: the one-video representation learning method DoRA, standard VideoMAE on multi-video datasets, and the task of future video prediction. To address this drop, we propose a geometric modification to standard optimizers, to decorrelate batches by utilising orthogonal gradients during training. The proposed modification can be applied to any optimizer -- we demonstrate it with Stochastic Gradient Descent (SGD) and AdamW. Our proposed orthogonal optimizer allows models trained from streaming videos to alleviate the drop in representation learning performance, as evaluated on downstream tasks. On three scenarios (DoRA, VideoMAE, future prediction), we show our orthogonal optimizer outperforms the strong AdamW in all three scenarios.

Joseph Heyward, João Carreira, Dima Damen et al.

Following the successful 2023 edition, we organised the Second Perception Test challenge as a half-day workshop alongside the IEEE/CVF European Conference on Computer Vision (ECCV) 2024, with the goal of benchmarking state-of-the-art video models and measuring the progress since last year using the Perception Test benchmark. This year, the challenge had seven tracks (up from six last year) and covered low-level and high-level tasks, with language and non-language interfaces, across video, audio, and text modalities; the additional track covered hour-long video understanding and introduced a novel video QA benchmark 1h-walk VQA. Overall, the tasks in the different tracks were: object tracking, point tracking, temporal action localisation, temporal sound localisation, multiple-choice video question-answering, grounded video question-answering, and hour-long video question-answering. We summarise in this report the challenge tasks and results, and introduce in detail the novel hour-long video QA benchmark 1h-walk VQA.

Karsten Roth, Zeynep Akata, Dima Damen et al.

Large-scale multimodal representation learning successfully optimizes for zero-shot transfer at test time. Yet the standard pretraining paradigm (contrastive learning on large amounts of image-text data) does not explicitly encourage representations to support few-shot adaptation. In this work, we propose a simple, but carefully designed extension to multimodal pretraining which enables representations to accommodate additional context. Using this objective, we show that vision-language models can be trained to exhibit significantly increased few-shot adaptation: across 21 downstream tasks, we find up to four-fold improvements in test-time sample efficiency, and average few-shot adaptation gains of over 5%, while retaining zero-shot generalization performance across model scales and training durations. In particular, equipped with simple, training-free, metric-based adaptation mechanisms, our representations easily surpass more complex and expensive optimization-based schemes, vastly simplifying generalization to new domains.

Toby Perrett, Tengda Han, Dima Damen et al.

Long videos contain many repeating actions, events and shots. These repetitions are frequently given identical captions, which makes it difficult to retrieve the exact desired clip using a text search. In this paper, we formulate the problem of unique captioning: Given multiple clips with the same caption, we generate a new caption for each clip that uniquely identifies it. We propose Captioning by Discriminative Prompting (CDP), which predicts a property that can separate identically captioned clips, and use it to generate unique captions. We introduce two benchmarks for unique captioning, based on egocentric footage and timeloop movies - where repeating actions are common. We demonstrate that captions generated by CDP improve text-to-video R@1 by 15% for egocentric videos and 10% in timeloop movies.

Masashi Hatano, Zhifan Zhu, Hideo Saito et al.

Forecasting hand motion and pose from an egocentric perspective is essential for understanding human intention. However, existing methods focus solely on predicting positions without considering articulation, and only when the hands are visible in the field of view. This limitation overlooks the fact that approximate hand positions can still be inferred even when they are outside the camera's view. In this paper, we propose a method to forecast the 3D trajectories and poses of both hands from an egocentric video, both in and out of the field of view. We propose a diffusion-based transformer architecture for Egocentric Hand Forecasting, EgoH4, which takes as input the observation sequence and camera poses, then predicts future 3D motion and poses for both hands of the camera wearer. We leverage full-body pose information, allowing other joints to provide constraints on hand motion. We denoise the hand and body joints along with a visibility predictor for hand joints and a 3D-to-2D reprojection loss that minimizes the error when hands are in-view. We evaluate EgoH4 on the Ego-Exo4D dataset, combining subsets with body and hand annotations. We train on 156K sequences and evaluate on 34K sequences, respectively. EgoH4 improves the performance by 3.4cm and 5.1cm over the baseline in terms of ADE for hand trajectory forecasting and MPJPE for hand pose forecasting. Project page: https://masashi-hatano.github.io/EgoH4/

Adriano Fragomeni, Dima Damen, Michael Wray

Video retrieval requires aligning visual content with corresponding natural language descriptions. In this paper, we introduce Modality Auxiliary Concepts for Video Retrieval (MAC-VR), a novel approach that leverages modality-specific tags -- automatically extracted from foundation models -- to enhance video retrieval. We propose to align modalities in a latent space, along with learning and aligning auxiliary latent concepts derived from the features of a video and its corresponding caption. We introduce these auxiliary concepts to improve the alignment of visual and textual latent concepts, allowing concepts to be distinguished from one another. We conduct extensive experiments on six diverse datasets: two different splits of MSR-VTT, DiDeMo, TGIF, Charades and YouCook2. The experimental results consistently demonstrate that modality-specific tags improve cross-modal alignment, outperforming current state-of-the-art methods across three datasets and performing comparably or better across others. Project Webpage: https://adrianofragomeni.github.io/MAC-VR/

Bin Zhu, Kevin Flanagan, Adriano Fragomeni et al.

Though pre-training vision-language models have demonstrated significant benefits in boosting video-text retrieval performance from large-scale web videos, fine-tuning still plays a critical role with manually annotated clips with start and end times, which requires considerable human effort. To address this issue, we explore an alternative cheaper source of annotations, single timestamps, for video-text retrieval. We initialise clips from timestamps in a heuristic way to warm up a retrieval model. Then a video clip editing method is proposed to refine the initial rough boundaries to improve retrieval performance. A student-teacher network is introduced for video clip editing. The teacher model is employed to edit the clips in the training set whereas the student model trains on the edited clips. The teacher weights are updated from the student's after the student's performance increases. Our method is model agnostic and applicable to any retrieval models. We conduct experiments based on three state-of-the-art retrieval models, COOT, VideoCLIP and CLIP4Clip. Experiments conducted on three video retrieval datasets, YouCook2, DiDeMo and ActivityNet-Captions show that our edited clips consistently improve retrieval performance over initial clips across all the three retrieval models.