The goal of this paper is to discover, segment, and track independently moving objects in complex visual scenes. Previous approaches have explored the use of optical flow for motion segmentation, leading to imperfect predictions due to partial motion, background distraction, and object articulations and interactions. To address this issue, we introduce an appearance-based refinement method that leverages temporal consistency in video streams to correct inaccurate flow-based proposals. Our approach involves a simple selection mechanism that identifies accurate flow-predicted masks as exemplars, and an object-centric architecture that refines problematic masks based on exemplar information. The model is pre-trained on synthetic data and then adapted to real-world videos in a self-supervised manner, eliminating the need for human annotations. Its performance is evaluated on multiple video segmentation benchmarks, including DAVIS, YouTubeVOS, SegTrackv2, and FBMS-59. We achieve competitive performance on single-object segmentation, while significantly outperforming existing models on the more challenging problem of multi-object segmentation. Finally, we investigate the benefits of using our model as a prompt for a per-frame Segment Anything Model.
Understanding long, real-world videos requires modeling of long-range visual dependencies. To this end, we explore video-first architectures, building on the common paradigm of transferring large-scale, image--text models to video via shallow temporal fusion. However, we expose two limitations to the approach: (1) decreased spatial capabilities, likely due to poor video--language alignment in standard video datasets, and (2) higher memory consumption, bottlenecking the number of frames that can be processed. To mitigate the memory bottleneck, we systematically analyze the memory/accuracy trade-off of various efficient methods: factorized attention, parameter-efficient image-to-video adaptation, input masking, and multi-resolution patchification. Surprisingly, simply masking large portions of the video (up to 75%) during contrastive pre-training proves to be one of the most robust ways to scale encoders to videos up to 4.3 minutes at 1 FPS. Our simple approach for training long video-to-text models, which scales to 1B parameters, does not add new architectural complexity and is able to outperform the popular paradigm of using much larger LLMs as an information aggregator over segment-based information on benchmarks with long-range temporal dependencies (YouCook2, EgoSchema).
We introduce a framework for online learning from a single continuous video stream -- the way people and animals learn, without mini-batches, data augmentation or shuffling. This poses great challenges given the high correlation between consecutive video frames and there is very little prior work on it. Our framework allows us to do a first deep dive into the topic and includes a collection of streams and tasks composed from two existing video datasets, plus methodology for performance evaluation that considers both adaptation and generalization. We employ pixel-to-pixel modelling as a practical and flexible way to switch between pre-training and single-stream evaluation as well as between arbitrary tasks, without ever requiring changes to models and always using the same pixel loss. Equipped with this framework we obtained large single-stream learning gains from pre-training with a novel family of future prediction tasks, found that momentum hurts, and that the pace of weight updates matters. The combination of these insights leads to matching the performance of IID learning with batch size 1, when using the same architecture and without costly replay buffers.
In this paper we tackle the problem of Generalized Category Discovery (GCD). Specifically, given a dataset with labelled and unlabelled images, the task is to cluster all images in the unlabelled subset, whether or not they belong to the labelled categories. Our first contribution is to recognize that most existing GCD benchmarks only contain labels for a single clustering of the data, making it difficult to ascertain whether models are using the available labels to solve the GCD task, or simply solving an unsupervised clustering problem. As such, we present a synthetic dataset, named 'Clevr-4', for category discovery. Clevr-4 contains four equally valid partitions of the data, i.e based on object shape, texture, color or count. To solve the task, models are required to extrapolate the taxonomy specified by the labelled set, rather than simply latching onto a single natural grouping of the data. We use this dataset to demonstrate the limitations of unsupervised clustering in the GCD setting, showing that even very strong unsupervised models fail on Clevr-4. We further use Clevr-4 to examine the weaknesses of existing GCD algorithms, and propose a new method which addresses these shortcomings, leveraging consistent findings from the representation learning literature to do so. Our simple solution, which is based on 'mean teachers' and termed $\mu$GCD, substantially outperforms implemented baselines on Clevr-4. Finally, when we transfer these findings to real data on the challenging Semantic Shift Benchmark (SSB), we find that $\mu$GCD outperforms all prior work, setting a new state-of-the-art. For the project webpage, see https://www.robots.ox.ac.uk/~vgg/data/clevr4/
Our objective in this paper is to estimate spine curvature in DXA scans. To this end we first train a neural network to predict the middle spine curve in the scan, and then use an integral-based method to determine the curvature along the spine curve. We use the curvature to compare to the standard angle scoliosis measure obtained using the DXA Scoliosis Method (DSM). The performance improves over the prior work of Jamaludin et al. 2018. We show that the maximum curvature can be used as a scoring function for ordering the severity of spinal deformation.
Auditory and visual signals usually present together and correlate with each other, not only in natural environments but also in clinical settings. However, the audio-visual modelling in the latter case can be more challenging, due to the different sources of audio/video signals and the noise (both signal-level and semantic-level) in auditory signals -- usually speech. In this paper, we consider audio-visual modelling in a clinical setting, providing a solution to learn medical representations that benefit various clinical tasks, without human expert annotation. A simple yet effective multi-modal self-supervised learning framework is proposed for this purpose. The proposed approach is able to localise anatomical regions of interest during ultrasound imaging, with only speech audio as a reference. Experimental evaluations on a large-scale clinical multi-modal ultrasound video dataset show that the proposed self-supervised method learns good transferable anatomical representations that boost the performance of automated downstream clinical tasks, even outperforming fully-supervised solutions.
Audio Description (AD) is the task of generating descriptions of visual content, at suitable time intervals, for the benefit of visually impaired audiences. For movies, this presents notable challenges -- AD must occur only during existing pauses in dialogue, should refer to characters by name, and ought to aid understanding of the storyline as a whole. To this end, we develop a new model for automatically generating movie AD, given CLIP visual features of the frames, the cast list, and the temporal locations of the speech; addressing all three of the 'who', 'when', and 'what' questions: (i) who -- we introduce a character bank consisting of the character's name, the actor that played the part, and a CLIP feature of their face, for the principal cast of each movie, and demonstrate how this can be used to improve naming in the generated AD; (ii) when -- we investigate several models for determining whether an AD should be generated for a time interval or not, based on the visual content of the interval and its neighbours; and (iii) what -- we implement a new vision-language model for this task, that can ingest the proposals from the character bank, whilst conditioning on the visual features using cross-attention, and demonstrate how this improves over previous architectures for AD text generation in an apples-to-apples comparison.
Recent advances in generative models like Stable Diffusion enable the generation of highly photo-realistic images. Our objective in this paper is to probe the diffusion network to determine to what extent it 'understands' different properties of the 3D scene depicted in an image. To this end, we make the following contributions: (i) We introduce a protocol to evaluate whether a network models a number of physical 'properties' of the 3D scene by probing for explicit features that represent these properties. The probes are applied on datasets of real images with annotations for the property. (ii) We apply this protocol to properties covering scene geometry, scene material, support relations, lighting, and view dependent measures. (iii) We find that Stable Diffusion is good at a number of properties including scene geometry, support relations, shadows and depth, but less performant for occlusion. (iv) We also apply the probes to other models trained at large-scale, including DINO and CLIP, and find their performance inferior to that of Stable Diffusion.
In this paper we introduce a new synchronisation task, Gesture-Sync: determining if a person's gestures are correlated with their speech or not. In comparison to Lip-Sync, Gesture-Sync is far more challenging as there is a far looser relationship between the voice and body movement than there is between voice and lip motion. We introduce a dual-encoder model for this task, and compare a number of input representations including RGB frames, keypoint images, and keypoint vectors, assessing their performance and advantages. We show that the model can be trained using self-supervised learning alone, and evaluate its performance on the LRS3 dataset. Finally, we demonstrate applications of Gesture-Sync for audio-visual synchronisation, and in determining who is the speaker in a crowd, without seeing their faces. The code, datasets and pre-trained models can be found at: \url{https://www.robots.ox.ac.uk/~vgg/research/gestsync}.
The goal of this paper is to detect what has changed, if anything, between two "in the wild" images of the same 3D scene acquired from different camera positions and at different temporal instances. The open-set nature of this problem, occlusions/dis-occlusions due to the shift in viewpoint, and the lack of suitable training datasets, presents substantial challenges in devising a solution. To address this problem, we contribute a change detection model that is trained entirely on synthetic data and is class-agnostic, yet it is performant out-of-the-box on real world images without requiring fine-tuning. Our solution entails a "register and difference" approach that leverages self-supervised frozen embeddings and feature differences, which allows the model to generalise to a wide variety of scenes and domains. The model is able to operate directly on two RGB images, without requiring access to ground truth camera intrinsics, extrinsics, depth maps, point clouds, or additional before-after images. Finally, we collect and release a new evaluation dataset consisting of real-world image pairs with human-annotated differences and demonstrate the efficacy of our method. The code, datasets and pre-trained model can be found at: https://github.com/ragavsachdeva/CYWS-3D