Abstract Out-of-distribution recognition forms an important and well-studied problem in deep learning, with the goal to filter out samples that do not belong to the distribution on which a network has been trained. The conclusion of this paper is simple: a good hierarchical hyperbolic embedding is preferred for discriminating in- and out-of-distribution samples. We introduce Balanced Hyperbolic Learning. We outline a hyperbolic class embedding algorithm that jointly optimizes for hierarchical distortion and balancing between shallow and wide subhierarchies. We then use the class embeddings as hyperbolic prototypes for classification on in-distribution data. We outline how to generalize existing out-of-distribution scoring functions to operate with hyperbolic prototypes. Empirical evaluations across 13 datasets and 13 scoring functions show that our hyperbolic embeddings outperform existing out-of-distribution approaches when trained on the same data with the same backbones. We also show that our hyperbolic embeddings outperform other hyperbolic approaches, beat state-of-the-art contrastive methods, and natively enable hierarchical out-of-distribution generalization.

Abstract The automatic understanding of video content is advancing rapidly. Empowered by deeper neural networks and large datasets, machines are increasingly capable of understanding what is concretely visible in video frames, whether it be objects, actions, events, or scenes. In comparison, humans retain a unique ability to also look beyond concrete entities and recognize abstract concepts like justice, freedom, and togetherness. Abstract concept recognition forms a crucial open challenge in video understanding, where reasoning on multiple semantic levels based on contextual information is key. In this paper, we argue that the recent advances in Foundation Models make for an ideal setting to address abstract concept understanding in videos. Automated understanding of high-level abstract concepts is imperative as it enables models to be more aligned with human reasoning and values. In this survey, we study different tasks and datasets used to understand abstract concepts in video content. We observe that, periodically and over a long period, researchers have attempted to solve these tasks, making the best use of the tools available at their disposal. We advocate that drawing on decades of community experience will help us shed light on this important open grand challenge and avoid “re-inventing the wheel” as we start revisiting it in the era of multi-modal Foundation Models.

Addressing the retrieval of unsafe content from vision-language models such as CLIP is an important step towards real-world integration. Current efforts have relied on unlearning techniques that try to erase the model’s knowledge of unsafe concepts. While effective in reducing unwanted outputs, unlearning limits the model’s capacity to discern between safe and unsafe content. In this work, we introduce a novel approach that shifts from unlearning to an awareness paradigm by leveraging the inherent hierarchical properties of the hyperbolic space. We propose to encode safe and unsafe content as an entailment hierarchy, where both are placed in different regions of hyperbolic space. Our HySAC, Hyperbolic Safety-Aware CLIP, employs entailment loss functions to model the hierarchical and asymmetrical relations between safe and unsafe image-text pairs. This modelling – ineffective in standard vision-language models due to their reliance on Euclidean embeddings – endows the model with awareness of unsafe content, enabling it to serve as both a multimodal unsafe classifier and a flexible content retriever, with the option to dynamically redirect unsafe queries toward safer alternatives or retain the original output. Extensive experiments show that our approach not only enhances safety recognition but also establishes a more adaptable and interpretable framework for content moderation in vision-language models. Our source code is available at: https://github.com/aimagelab/HySAC

This paper revisits the problem of predicting box locations in object detection architectures. Typically, each box proposal or box query aims to directly maximize the intersection-over-union score with the ground truth, followed by a winner-takes-all non-maximum suppression where only the highest scoring box in each region is retained. We observe that both steps are sub-optimal: the first involves regressing proposals to the entire ground truth, which is a difficult task even with large receptive fields, and the second neglects valuable information from boxes other than the top candidate. Instead of regressing proposals to the whole ground truth, we propose a simpler approach—regress only to the area of intersection between the proposal and the ground truth. This avoids the need for proposals to extrapolate beyond their visual scope, improving localization accuracy. Rather than adopting a winner-takes-all strategy, we take the union over the regressed intersections of all boxes in a region to generate the final box outputs. Our plug-and-play method integrates seamlessly into proposal-based, grid-based, and query-based detection architectures with minimal modifications, consistently improving object localization and instance segmentation. We demonstrate its broad applicability and versatility across various detection and segmentation tasks.

The human brain efficiently represents visual inputs through specialized neural populations that selectively respond to specific categories. Advancements in generative modeling have enabled data-driven discovery of neural selectivity using brain-optimized image synthesis. However, current methods independently generate one sample at a time, without enforcing structural constraints on the generations; thus, these individual images have no explicit point of comparison, making it hard to discern which image features drive neural response selectivity. To address this issue, we introduce Brain Activation Control Through Image Variation (BrainACTIV), a method for manipulating a reference image to enhance or decrease activity in a target cortical region using pretrained diffusion models. Starting from a reference image allows for fine-grained and reliable offline identification of optimal visuo-semantic properties, as well as producing controlled stimuli for novel neuroimaging studies. We show that our manipulations effectively modulate predicted fMRI responses and agree with hypothesized preferred categories in established regions of interest, while remaining structurally close to the reference image. Moreover, we demonstrate how our method accentuates differences between brain regions that are selective to the same category, and how it could be used to explore neural representation of brain regions with unknown selectivities. Hence, BrainACTIV holds the potential to formulate robust hypotheses about brain representation and to facilitate the production of naturalistic stimuli for neuroscientific experiments.

Image-text representation learning forms a cornerstone in vision-language models, where pairs of images and textual descriptions are contrastively aligned in a shared embedding space. Since visual and textual concepts are naturally hierarchical, recent work has shown that hyperbolic space can serve as a high-potential manifold to learn vision-language representation with strong downstream performance. In this work, for the first time we show how to fully leverage the innate hierarchical nature of hyperbolic embeddings by looking beyond individual image-text pairs. We propose Compositional Entailment Learning for hyperbolic vision-language models. The idea is that an image is not only described by a sentence but is itself a composition of multiple object boxes, each with their own textual description. Such information can be obtained freely by extracting nouns from sentences and using openly available localized grounding models. We show how to hierarchically organize images, image boxes, and their textual descriptions through contrastive and entailment-based objectives. Empirical evaluation on a hyperbolic vision-language model trained with millions of image-text pairs shows that the proposed compositional learning approach outperforms conventional Euclidean CLIP learning, as well as recent hyperbolic alternatives, with better zero-shot and retrieval generalization and clearly stronger hierarchical performance.

Embedding tree-like data, from hierarchies to ontologies and taxonomies, forms a well-studied problem for representing knowledge across many domains. Hyperbolic geometry provides a natural solution for embedding trees, with vastly superior performance over Euclidean embeddings. Recent literature has shown that hyperbolic tree embeddings can even be placed on top of neural networks for hierarchical knowledge integration in deep learning settings. For all applications, a faithful embedding of trees is needed, with combinatorial constructions emerging as the most effective direction. This paper identifies and solves two key limitations of existing works. First, the combinatorial construction hinges on finding highly separated points on a hypersphere, a notoriously difficult problem. Current approaches achieve poor separation, degrading the quality of the corresponding hyperbolic embedding. We propose highly separated Delaunay tree embeddings (HS-DTE), which integrates angular separation in a generalized formulation of Delaunay embeddings, leading to lower embedding distortion. Second, low-distortion requires additional precision. The current approach for increasing precision is to use multiple precision arithmetic, which renders the embeddings useless on GPUs in deep learning settings. We reformulate the combinatorial construction using floating point expansion arithmetic, leading to superior embedding quality while retaining utility on accelerated hardware.

Hyperbolic geometry is an effective geometry for embedding hierarchical data structures. Hyperbolic learning has therefore become increasingly prominent in machine learning applications where data is hierarchically organized or governed by hierarchical semantics, ranging from recommendation systems to computer vision. The quality of hyperbolic embeddings is tightly coupled to the structure of the input hierarchy, which is often derived from knowledge graphs or ontologies. Recent work has uncovered that for an optimal hyperbolic embedding, a high branching factor and single inheritance are key, while embedding algorithms are robust to imbalance and hierarchy size. To assist knowledge engineers in reorganizing hierarchical knowledge, this paper investigates whether Large Language Models (LLMs) have the ability to automatically restructure hierarchies to meet these criteria. We propose a prompt-based approach to transform existing hierarchies using LLMs, guided by known desiderata for hyperbolic embeddings. Experiments on 16 diverse hierarchies show that LLM-restructured hierarchies consistently yield higher-quality hyperbolic embeddings across several standard embedding quality metrics. Moreover, we show how LLM-guided hierarchy restructuring enables explainable reorganizations, providing justifications to knowledge engineers.

Abstract Zero-shot recognition is centered around learning representations to transfer knowledge from seen to unseen classes. Where foundational approaches perform the transfer with semantic embedding spaces, e.g., from attributes or word vectors, the current state-of-the-art relies on prompting pre-trained vision-language models to obtain class embeddings. Whether zero-shot learning is performed with attributes, CLIP, or something else, current approaches de facto assume that there is a pre-defined embedding space in which seen and unseen classes can be positioned. Our work is concerned with real-world zero-shot settings where a pre-defined embedding space can no longer be assumed. This is natural in domains such as biology and medicine, where class names are not common English words, rendering vision-language models useless; or neuroscience, where class relations are only given with non-semantic human comparison scores. We find that there is one data structure enabling zero-shot learning in both standard and non-standard settings: a similarity matrix spanning the seen and unseen classes. We introduce four similarity-based zero-shot learning challenges, tackling open-ended scenarios such as learning with uncommon class names, learning from multiple partial sources, and learning with missing knowledge. As the first step for zero-shot learning beyond a pre-defined semantic embedding space, we propose $$\kappa $$ κ -MDS, a general approach that obtains a prototype for each class on any manifold from similarities alone, even when part of the similarities are missing. Our approach can be plugged into any standard, hyperspherical, or hyperbolic zero-shot learner. Experiments on existing datasets and the new benchmarks show the promise and challenges of similarity-based zero-shot learning.

This paper strives for image editing via generative models. Flow Matching is an emerging generative modeling technique that offers the advantage of simple and efficient training. Simultaneously, a new transformer-based U-ViT has recently been proposed to replace the commonly used UNet for better scalability and performance in generative modeling. Hence, Flow Matching with a transformer backbone offers the potential for scalable and high-quality generative modeling, but their latent structure and editing ability are as of yet unknown. Hence, we adopt this setting and explore how to edit images through latent space manipulation. We introduce an editing space, which we call u-space, that can be manipulated in a controllable, accumulative, and composable manner. Additionally, we propose a tailored sampling solution to enable sampling with the more efficient adaptive step-size ODE solvers. Lastly, we put forth a straightforward yet powerful method for achieving fine-grained and nuanced editing using text prompts. Our framework is simple and efficient, all while being highly effective at editing images while preserving the essence of the original content. Our code will be publicly available at https://taohu.me/lfm/

Autonomous robots are increasingly becoming a strong fixture in social environments. Effective crowd navigation requires not only safe yet fast planning, but should also enable interpretability and computational efficiency for working in real-time on embedded devices. In this work, we advocate for hyperbolic learning to enable crowd navigation and we introduce Hyp2Nav. Different from conventional reinforcement learning-based crowd navigation methods, Hyp2Nav leverages the intrinsic properties of hyperbolic geometry to better encode the hierarchical nature of decision-making processes in navigation tasks. We propose a hyperbolic policy model and a hyperbolic curiosity module that results in effective social navigation, best success rates, and returns across multiple simulation settings, using up to 6 times fewer parameters than competitor state-of-the-art models. With our approach, it becomes even possible to obtain policies that work in 2-dimensional embedding spaces, opening up new possibilities for low-resource crowd navigation and model interpretability. Insightfully, the internal hyperbolic representation of Hyp2Nav correlates with how much attention the robot pays to the surrounding crowds, e.g. due to multiple people occluding its pathway or to a few of them showing colliding plans, rather than to its own planned route. The code is available at https://github.com/GDam90/hyp2nav.

Abstract Deep representation learning is a ubiquitous part of modern computer vision. While Euclidean space has been the de facto standard manifold for learning visual representations, hyperbolic space has recently gained rapid traction for learning in computer vision. Specifically, hyperbolic learning has shown a strong potential to embed hierarchical structures, learn from limited samples, quantify uncertainty, add robustness, limit error severity, and more. In this paper, we provide a categorization and in-depth overview of current literature on hyperbolic learning for computer vision. We research both supervised and unsupervised literature and identify three main research themes in each direction. We outline how hyperbolic learning is performed in all themes and discuss the main research problems that benefit from current advances in hyperbolic learning for computer vision. Moreover, we provide a high-level intuition behind hyperbolic geometry and outline open research questions to further advance research in this direction.

This paper introduces an end-to-end residual network that operates entirely on the Poincaré ball model of hyperbolic space. Hyperbolic learning has recently shown great potential for visual understanding, but is currently only performed in the penultimate layer(s) of deep networks. All visual representations are still learned through standard Euclidean networks. In this paper we investigate how to learn hyperbolic representations of visual data directly from the pixel-level. We propose Poincaré ResNet, a hyperbolic counterpart of the celebrated residual network, starting from Poincaré 2D convolutions up to Poincaré residual connections. We identify three roadblocks for training convolutional networks entirely in hyperbolic space and propose a solution for each: (i) Current hyperbolic network initializations collapse to the origin, limiting their applicability in deeper networks. We provide an identity-based initialization that preserves norms over many layers. (ii) Residual networks rely heavily on batch normalization, which comes with expensive Fréchet mean calculations in hyperbolic space. We introduce Poincaré midpoint batch normalization as a faster and equally effective alternative. (iii) Due to the many intermediate operations in Poincaré layers, the computation graphs of deep learning libraries blow up, limiting our ability to train on deep hyperbolic networks. We provide manual backward derivations of core hyperbolic operations to maintain manageable computation graphs.

This paper considers retrieval of videos containing human activity from just a video query. In the literature, a common assumption is that all activities have sufficient labelled examples when learning an embedding for retrieval. However, this assumption does not hold in practice, as only a portion of activities have many examples, while other activities are only described by few examples. In this paper, we propose a visual-semantic embedding network that explicitly deals with the imbalanced scenario for activity retrieval. Our network contains two novel modules. The visual alignment module performs a global alignment between the input video and visual feature bank representations for all activities. The semantic module performs an alignment between the input video and semantic activity representations. By matching videos with both visual and semantic activity representations over all activities, we no longer ignore infrequent activities during retrieval. Experiments on a new imbalanced activity retrieval benchmark show the effectiveness of our proposal.

This paper investigates the problem of scene graph generation in videos with the aim of capturing semantic relations between subjects and objects in the form of ⟨ subject, predicate, object⟩ triplets. Recognizing the predicate between subject and object pairs is imbalanced and multi-label in nature, ranging from ubiquitous interactions such as spatial relationships (e.g. in front of) to rare interactions such as twisting. In widely-used benchmarks such as Action Genome and VidOR, the imbalance ratio between the most and least frequent predicates reaches 3,218 and 3,408, respectively, surpassing even benchmarks specifically designed for long-tailed recognition. Due to the long-tailed distributions and label co-occurrences, recent state-of-the-art methods predominantly focus on the most frequently occurring predicate classes, ignoring those in the long tail. In this paper, we analyze the limitations of current approaches for scene graph generation in videos and identify a one-to-one correspondence between predicate frequency and recall performance. To make the step towards unbiased scene graph generation in videos, we introduce a multi-label meta-learning framework to deal with the biased predicate distribution. Our meta-learning framework learns a meta-weight network for each training sample over all possible label losses. We evaluate our approach on the Action Genome and VidOR benchmarks by building upon two current state-of-the-art methods for each benchmark. The experiments demonstrate that the multi-label meta-weight network improves the performance for predicates in the long tail without compromising performance for head classes, resulting in better overall performance and favorable generalizability. Code: https://github.com/shanshuo/ML-MWN.

Deep learning in hyperbolic space is quickly gaining traction in the fields of machine learning, multimedia, and computer vision. Deep networks commonly operate in Euclidean space, implicitly assuming that data lies on regular grids. Recent advances have shown that hyperbolic geometry provides a viable alternative foundation for deep learning, especially when data is hierarchical in nature and when working with few embedding dimensions. Currently however, no accessible open-source library exists to build hyperbolic network modules akin to well-known deep learning libraries. We present HypLL, the Hyperbolic Learning Library to bring the progress on hyperbolic deep learning together. HypLL is built on top of PyTorch, with an emphasis in its design for ease-of-use, in order to attract a broad audience towards this new and open-ended research direction. The code is available at: https://github.com/maxvanspengler/hyperbolic_learning_library.

Abstract This work addresses the problem of recognizing action categories in videos when no training examples are available. The current state-of-the-art enables such a zero-shot recognition by learning universal mappings from videos to a semantic space, either trained on large-scale seen actions or on objects. While effective, we find that universal action and object mappings are biased to specific regions in the semantic space. These biases lead to a fundamental problem: many unseen action categories are simply never inferred during testing. For example on UCF-101, a quarter of the unseen actions are out of reach with a state-of-the-art universal action model. To that end, this paper introduces universal prototype transport for zero-shot action recognition. The main idea is to re-position the semantic prototypes of unseen actions by matching them to the distribution of all test videos. For universal action models, we propose to match distributions through a hyperspherical optimal transport from unseen action prototypes to the set of all projected test videos. The resulting transport couplings in turn determine the target prototype for each unseen action. Rather than directly using the target prototype as final result, we re-position unseen action prototypes along the geodesic spanned by the original and target prototypes as a form of semantic regularization. For universal object models, we outline a variant that defines target prototypes based on an optimal transport between unseen action prototypes and object prototypes. Empirically, we show that universal prototype transport diminishes the biased selection of unseen action prototypes and boosts both universal action and object models for zero-shot classification and spatio-temporal localization.

This article introduces the task of visual named entity discovery in videos without the need for task-specific supervision or task-specific external knowledge sources. Assigning specific names to entities (e.g., faces, scenes, or objects) in video frames is a long-standing challenge. Commonly, this problem is addressed as a supervised learning objective by manually annotating entities with labels. To bypass the annotation burden of this setup, several works have investigated the problem by utilizing external knowledge sources such as movie databases. While effective, such approaches do not work when task-specific knowledge sources are not provided and can only be applied to movies and TV series. In this work, we take the problem a step further and propose to discover entities in videos from videos and corresponding captions or subtitles. We introduce a three-stage method where we (i) create bipartite entity-name graphs from frame–caption pairs, (ii) find visual entity agreements, and (iii) refine the entity assignment through entity-level prototype construction. To tackle this new problem, we outline two new benchmarks, SC-Friends and SC-BBT , based on the Friends and Big Bang Theory TV series. Experiments on the benchmarks demonstrate the ability of our approach to discover which named entity belongs to which face or scene, with an accuracy close to a supervised oracle, just from the multimodal information present in videos. Additionally, our qualitative examples show the potential challenges of self-contained discovery of any visual entity for future work. The code and the data are available on GitHub. 1

For image segmentation, the current standard is to perform pixel-level optimization and inference in Euclidean output embedding spaces through linear hyperplanes. In this work, we show that hyperbolic manifolds provide a valuable alternative for image segmentation and propose a tractable formulation of hierarchical pixel-level classification in hyperbolic space. Hyperbolic Image Segmentation opens up new possibilities and practical benefits for segmentation, such as uncertainty estimation and boundary information for free, zero-label generalization, and increased performance in low-dimensional output embeddings.

This paper focuses on self-supervised representation learning in videos with guidance from multimodal priors. Where the temporal dimension is commonly used as supervision proxy for learning frame-level or clip-level representations, a number of works have recently shown how to learn local representations in space and time through cycle-consistency. Given a starting patch, the contrastive goal is to track the patch in subsequent frames, followed by a backtracking to the original frame with the starting patch as goal. While effective for down-stream tasks such as segmentation and body joint propagation, affinities between patches need to be learned from scratch. This setup not only requires many videos for self-supervised optimization, it also fails when using smaller patches and more connections between consecutive frames. On the other hand, there are multiple generic cues from multiple modalities that provide valuable information about how patches should propagate in videos, from saliency and optical flow to photometric center biases. To that end, we introduce Guided Contrastive Random Walks. The main idea is to employ well-known multimodal priors to provide fixed prior affinities. We outline a general framework where prior affinities are combined with learned affinities to guide the cycle-consistency objective. Empirically, we show that Guided Contrastive Random Walks result in better spatio-temporal representations for two down-stream tasks. More importantly, when using smaller patches and therefore more connections between patches, our approach further improves, while the unguided baseline can no longer learn meaningful representations.

Maximizing the separation between classes constitutes a well-known inductive bias in machine learning and a pillar of many traditional algorithms. By default, deep networks are not equipped with this inductive bias and therefore many alternative solutions have been proposed through differential optimization. Current approaches tend to optimize classification and separation jointly: aligning inputs with class vectors and separating class vectors angularly. This paper proposes a simple alternative: encoding maximum separation as an inductive bias in the network by adding one fixed matrix multiplication before computing the softmax activations. The main observation behind our approach is that separation does not require optimization but can be solved in closed-form prior to training and plugged into a network. We outline a recursive approach to obtain the matrix consisting of maximally separable vectors for any number of classes, which can be added with negligible engineering effort and computational overhead. Despite its simple nature, this one matrix multiplication provides real impact. We show that our proposal directly boosts classification, long-tailed recognition, out-of-distribution detection, and open-set recognition, from CIFAR to ImageNet. We find empirically that maximum separation works best as a fixed bias; making the matrix learnable adds nothing to the performance. The closed-form implementation and code to reproduce the experiments are available on github.

3D convolutional networks, as direct inheritors of 2D convolutional networks for images, have placed their mark on action recognition in videos. Combined with pretraining on large-scale video data, high classification accuracies have been obtained on numerous video benchmarks. In an effort to better understand why 3D convolutional networks are so effective, several works have highlighted their bias towards static appearance and towards the scenes in which actions occur. In this work, we seek to find the source of this bias and question whether the observed biases towards static appearances are inherent to 3D convolutional networks or represent limited significance of motion in the training data. We resolve this by presenting temporality measures that estimate the data-to-model motion dependency at both the layer-level and the kernel-level. Moreover, we introduce two synthetic datasets where motion and appearance are decoupled by design, which allows us to directly observe their effects on the networks. Our analysis shows that 3D architectures are not inherently biased towards appearance. When trained on the most prevalent video sets, 3D convolutional networks are indeed biased throughout, especially in the final layers of the network. However, when training on data with motions and appearances explicitly decoupled and balanced, such networks adapt to varying levels of temporality. To this end, we see the proposed measures as a reliable method to estimate motion relevance for activity classification in datasets and use them to uncover the differences between popular pre-training video collections, such as Kinetics, IG-65M and Howto100 m.

Abstract The deep image prior showed that a randomly initialized network with a suitable architecture can be trained to solve inverse imaging problems by simply optimizing it’s parameters to reconstruct a single degraded image. However, it suffers from two practical limitations. First, it remains unclear how to control the prior beyond the choice of the network architecture. Second, training requires an oracle stopping criterion as during the optimization the performance degrades after reaching an optimum value. To address these challenges we introduce a frequency-band correspondence measure to characterize the spectral bias of the deep image prior, where low-frequency image signals are learned faster and better than high-frequency counterparts. Based on our observations, we propose techniques to prevent the eventual performance degradation and accelerate convergence. We introduce a Lipschitz-controlled convolution layer and a Gaussian-controlled upsampling layer as plug-in replacements for layers used in the deep architectures. The experiments show that with these changes the performance does not degrade during optimization, relieving us from the need for an oracle stopping criterion. We further outline a stopping criterion to avoid superfluous computation. Finally, we show that our approach obtains favorable results compared to current approaches across various denoising, deblocking, inpainting, super-resolution and detail enhancement tasks. Code is available at https://github.com/shizenglin/Measure-and-Control-Spectral-Bias .

This paper introduces the task of few-shot common action localization in time and space. Given a few trimmed support videos containing the same but unknown action, we strive for spatio-temporal localization of that action in a long untrimmed query video. We do not require any class labels, interval bounds, or bounding boxes. To address this challenging task, we introduce a novel few-shot transformer architecture with a dedicated encoder-decoder structure optimized for joint commonality learning and localization prediction, without the need for proposals. Experiments on re-organizations of the AVA and UCF101-24 datasets show the effectiveness of our approach for few-shot common action localization, even when the support videos are noisy. Although we are not specifically designed for common localization in time only, we also compare favorably against the few-shot and one-shot state-of-the-art in this setting. Lastly, we demonstrate that the few-shot transformer is easily extended to common action localization per pixel.

This paper strives to classify and detect the relationship between object tubelets appearing within a video as a 〈subject-predicate-object〉 triplet. Where existing works treat object proposals or tubelets as single entities and model their relations a posteriori, we propose to classify and detect predicates for pairs of object tubelets a priori. We also propose Social Fabric: an encoding that represents a pair of object tubelets as a composition of interaction primitives. These primitives are learned over all relations, resulting in a compact representation able to localize and classify relations from the pool of co-occurring object tubelets across all timespans in a video. The encoding enables our two-stage network. In the first stage, we train Social Fabric to suggest proposals that are likely interacting. We use the Social Fabric in the second stage to simultaneously finetune and predict predicate labels for the tubelets. Experiments demonstrate the benefit of early video relation modeling, our encoding and the two-stage architecture, leading to a new state-of-the-art on two benchmarks. We also show how the encoding enables query-by-primitive-example to search for spatio-temporal video relations. Code: https://github.com/shanshuo/Social-Fabric.

Hyperbolic space has become a popular choice of manifold for representation learning of various datatypes from tree-like structures and text to graphs. Building on the success of deep learning with prototypes in Euclidean and hyperspherical spaces, a few recent works have proposed hyperbolic prototypes for classification. Such approaches enable effective learning in low-dimensional output spaces and can exploit hierarchical relations amongst classes, but require privileged information about class labels to position the hyperbolic prototypes. In this work, we propose Hyperbolic Busemann Learning. The main idea behind our approach is to position prototypes on the ideal boundary of the Poincar\'{e} ball, which does not require prior label knowledge. To be able to compute proximities to ideal prototypes, we introduce the penalised Busemann loss. We provide theory supporting the use of ideal prototypes and the proposed loss by proving its equivalence to logistic regression in the one-dimensional case. Empirically, we show that our approach provides a natural interpretation of classification confidence, while outperforming recent hyperspherical and hyperbolic prototype approaches.

Abstract This work strives for the classification and localization of human actions in videos, without the need for any labeled video training examples. Where existing work relies on transferring global attribute or object information from seen to unseen action videos, we seek to classify and spatio-temporally localize unseen actions in videos from image-based object information only. We propose three spatial object priors, which encode local person and object detectors along with their spatial relations. On top we introduce three semantic object priors, which extend semantic matching through word embeddings with three simple functions that tackle semantic ambiguity, object discrimination, and object naming. A video embedding combines the spatial and semantic object priors. It enables us to introduce a new video retrieval task that retrieves action tubes in video collections based on user-specified objects, spatial relations, and object size. Experimental evaluation on five action datasets shows the importance of spatial and semantic object priors for unseen actions. We find that persons and objects have preferred spatial relations that benefit unseen action localization, while using multiple languages and simple object filtering directly improves semantic matching, leading to state-of-the-art results for both unseen action classification and localization.

The goal of this paper is guided image filtering, which emphasizes the importance of structure transfer during filtering by means of an additional guidance image. Where classical guided filters transfer structures using hand-designed functions, recent guided filters have been considerably advanced through parametric learning of deep networks. The state-of-the-art leverages deep networks to estimate the two core coefficients of the guided filter. In this work, we posit that simultaneously estimating both coefficients is suboptimal, resulting in halo artifacts and structure inconsistencies. Inspired by unsharp masking, a classical technique for edge enhancement that requires only a single coefficient, we propose a new and simplified formulation of the guided filter. Our formulation enjoys a filtering prior from a low-pass filter and enables explicit structure transfer by estimating a single coefficient. Based on our proposed formulation, we introduce a successive guided filtering network, which provides multiple filtering results from a single network, allowing for a trade-off between accuracy and efficiency. Extensive ablations, comparisons and analysis show the effectiveness and efficiency of our formulation and network, resulting in state-of-the-art results across filtering tasks like upsampling, denoising, and cross-modality filtering. Code is available at https://github.com/shizenglin/Unsharp-Mask-Guided-Filtering.

This work proposes a metric learning approach for self-supervised scene flow estimation. Scene flow estimation is the task of estimating 3D flow vectors for consecutive 3D point clouds. Such flow vectors are fruitful, e.g. for recognizing actions, or avoiding collisions. Training a neural network via supervised learning for scene flow is impractical, as this requires manual annotations for each 3D point at each new timestamp for each scene. To that end, we seek for a self-supervised approach, where a network learns a latent metric to distinguish between points translated by flow estimations and the target point cloud. Our adversarial metric learning includes a multi-scale triplet loss on sequences of two-point clouds as well as a cycle consistency loss. Furthermore, we outline a benchmark for self-supervised scene flow estimation: the Scene Flow Sandbox. The benchmark consists of five datasets designed to study individual aspects of flow estimation in progressive order of complexity, from a moving object to real-world scenes. Experimental evaluation on the benchmark shows that our approach obtains state-of-the-art self-supervised scene flow results, outperforming recent neighbor-based approaches. We use our proposed benchmark to expose shortcomings and draw insights on various training setups. We find that our setup captures motion coherence and preserves local geometries. Dealing with occlusions, on the other hand, is still an open challenge.

In this paper, we introduce hierarchical action search. Starting from the observation that hierarchies are mostly ignored in the action literature, we retrieve not only individual actions but also relevant and related actions, given an action name or video example as input. We propose a hyperbolic action network, which is centered around a hyperbolic space shared by action hierarchies and videos. Our discriminative hyperbolic embedding projects actions on the shared space while jointly optimizing hypernym-hyponym relations between action pairs and a large margin separation between all actions. The projected actions serve as hyperbolic prototypes that we match with projected video representations. The result is a learned space where videos are positioned in entailment cones formed by different subtrees. To perform search in this space, we start from a query and increasingly enlarge its entailment cone to retrieve hierarchically relevant action videos. Experiments on three action datasets with new hierarchy annotations show the effectiveness of our approach for hierarchical action search by name and by video example, regardless of whether queried actions have been seen or not during training. Our implementation is available at https://github.com/Tenglon/hyperbolic_action.

In this work, we investigate activity recognition using multimodal inputs from heterogeneous sensors. Activity recognition is commonly tackled from a single-modal perspective using videos. In case multiple signals are used, they come from the same homogeneous modality, e.g. in the case of color and optical flow. Here, we propose an activity network that fuses multimodal inputs coming from completely different and heterogeneous sensors. We frame such a heterogeneous fusion as a non-local operation. The observation is that in a non-local operation, only the channel dimensions need to match. In the network, heterogeneous inputs are fused, while maintaining the shapes and dimensionalities that fit each input. We outline both asymmetric fusion, where one modality serves to enforce the other, and symmetric fusion variants. To further promote research into multimodal activity recognition, we introduce GloVid, a first-person activity dataset captured with video recordings and smart glove sensor readings. Experiments on GloVid show the potential of heterogeneous non-local fusion for activity recognition, outperforming individual modalities and standard fusion techniques.

This paper introduces spatio-temporal interactivity proposals for video surveillance. Rather than focusing solely on actions performed by subjects, we explicitly include the objects that the subjects interact with. To enable interactivity proposals, we introduce the notion of interactivityness, a score that reflects the likelihood that a subject and object have an interplay. For its estimation, we propose a network containing an interactivity block and geometric encoding between subjects and objects. The network computes local interactivity likelihoods from subject and object trajectories, which we use to link intervals of high scores into spatio-temporal proposals. Experiments on an interactivity dataset with new evaluation metrics show the general benefit of interactivity proposals as well as its favorable performance compared to traditional temporal and spatio-temporal action proposals.

This work investigates the anticipation of future ship locations based on multimodal sensors. Predicting future trajectories of ships is an important component for the development of safe autonomous sailing ships on water. A core challenge towards future trajectory prediction is making sense of multiple modalities from vastly different sensors, including GPS coordinates, radar images, and charts specifying water and land regions. To that end, we propose a Trajectory Prediction Network, an end-to-end approach for trajectory anticipation based on multimodal sensors. Our approach is framed as a multi-task sequence-to-sequence network, with network components for coordinate sequences and radar images. In the network, water/land segmentations from charts are integrated as an auxiliary training objective. Since future anticipation of ships has not previously been studied from such a multimodal perspective, we introduce the Inland Shipping Dataset (ISD), a novel dataset for future anticipation of ships. Experimental evaluation on ISD shows the potential of our approach, outperforming single-modal variants and baselines from related anticipation tasks.

This article aims for the detection and search of events in videos, where video examples are either scarce or even absent during training. To enable such event detection and search, ImageNet concept banks have shown to be effective. Rather than employing the standard concept bank of 1,000 ImageNet classes, we leverage the full 21,841-class dataset. We identify two problems with using the full dataset: (i) there is an imbalance between the number of examples per concept, and (ii) not all concepts are equally relevant for events. In this article, we propose to balance large-scale image hierarchies for pre-training. We shuffle concepts based on bottom-up and top-down operations to overcome the problems of example imbalance and concept relevance. Using this strategy, we arrive at the shuffled ImageNet bank, a concept bank with an order of magnitude more concepts compared to standard ImageNet banks. Compared to standard ImageNet pre-training, our shuffles result in more discriminative representations to train event models from the limited video event examples. For event search, the broad range of concepts enable a closer match between textual queries of events and concept detections in videos. Experimentally, we show the benefit of the proposed bank for event detection and event search, with state-of-the-art performance for both tasks on the challenging TRECVID Multimedia Event Detection and Ad-Hoc Video Search benchmarks.

Few-shot learning is a nascent research topic, motivated by the fact that traditional deep learning requires tremendous amounts of data. In this work, we propose a new task along this research direction, we call few-shot common-localization. Given a few weakly-supervised support images, we aim to localize the common object in the query image without any box annotation. This task differs from standard few-shot settings, since we aim to address the localization problem, rather than the global classification problem. To tackle this new problem, we propose a network that aims to get the most out of the support and query images. To that end, we introduce a spatial similarity module that searches the spatial commonality among the given images. We furthermore introduce a feature reweighting module to balance the influence of different support images through graph convolutional networks. To evaluate few-shot common-localization, we repurpose and reorganize the well-known Pascal VOC and MS-COCO datasets, as well as a video dataset from ImageNet VID. Experiments on the new settings for few-shot common-localization shows the importance of searching for spatial similarity and feature reweighting, outperforming baselines from related tasks.

This paper aims to count arbitrary objects in images. The leading counting approaches start from point annotations per object from which they construct density maps. Then, their training objective transforms input images to density maps through deep convolutional networks. We posit that the point annotations serve more supervision purposes than just constructing density maps. We introduce ways to repurpose the points for free. First, we propose supervised focus from segmentation, where points are converted into binary maps. The binary maps are combined with a network branch and accompanying loss function to focus on areas of interest. Second, we propose supervised focus from global density, where the ratio of point annotations to image pixels is used in another branch to regularize the overall density estimation. To assist both the density estimation and the focus from segmentation, we also introduce an improved kernel size estimator for the point annotations. Experiments on six datasets show that all our contributions reduce the counting error, regardless of the base network, resulting in state-of-the-art accuracy using only a single network. Finally, we are the first to count on WIDER FACE, allowing us to show the benefits of our approach in handling varying object scales and crowding levels. Code is available at https://github.com/shizenglin/Counting-with-Focus-for-Free.

This demo presents a system for journalists to explore video footage for broadcasts. Daily news broadcasts contain multiple news items that consist of many video shots and searching for relevant footage is a labor intensive task. Without the need for annotated video shots, our system extracts semantics from footage and automatically matches these semantics to query terms from the journalist. The journalist can then indicate which aspects of the query term need to be emphasized, e.g. the title or its thematic meaning. The goal of this system is to support the journalists in their search process by encouraging interaction and exploration with the system.

This paper introduces hyperspherical prototype networks, which unify classification and regression with prototypes on hyperspherical output spaces. For classification, a common approach is to define prototypes as the mean output vector over training examples per class. Here, we propose to use hyperspheres as output spaces, with class prototypes defined a priori with large margin separation. We position prototypes through data-independent optimization, with an extension to incorporate priors from class semantics. By doing so, we do not require any prototype updating, we can handle any training size, and the output dimensionality is no longer constrained to the number of classes. Furthermore, we generalize to regression, by optimizing outputs as an interpolation between two prototypes on the hypersphere. Since both tasks are now defined by the same loss function, they can be jointly trained for multi-task problems. Experimentally, we show the benefit of hyperspherical prototype networks for classification, regression, and their combination over other prototype methods, softmax cross-entropy, and mean squared error approaches.

This paper strives for spatio-temporal localization of human actions in videos. In the literature, the consensus is to achieve localization by training on bounding box annotations provided for each frame of each training video. As annotating boxes in video is expensive, cumbersome and error-prone, we propose to bypass box-supervision. Instead, we introduce action localization based on point-supervision. We start from unsupervised spatio-temporal proposals, which provide a set of candidate regions in videos. While normally used exclusively for inference, we show spatio-temporal proposals can also be leveraged during training when guided by a sparse set of point annotations. We introduce an overlap measure between points and spatio-temporal proposals and incorporate them all into a new objective of a multiple instance learning optimization. During inference, we introduce pseudo-points, visual cues from videos, that automatically guide the selection of spatio-temporal proposals. We outline five spatial and one temporal pseudo-point, as well as a measure to best leverage pseudo-points at test time. Experimental evaluation on three action localization datasets shows our pointly-supervised approach (1) is as effective as traditional box-supervision at a fraction of the annotation cost, (2) is robust to sparse and noisy point annotations, (3) benefits from pseudo-points during inference, and (4) outperforms recent weakly-supervised alternatives. This leads us to conclude that points provide a viable alternative to boxes for action localization.

We aim for zero-shot localization and classification of human actions in video. Where traditional approaches rely on global attribute or object classification scores for their zero-shot knowledge transfer, our main contribution is a spatial-aware object embedding. To arrive at spatial awareness, we build our embedding on top of freely available actor and object detectors. Relevance of objects is determined in a word embedding space and further enforced with estimated spatial preferences. Besides local object awareness, we also embed global object awareness into our embedding to maximize actor and object interaction. Finally, we exploit the object positions and sizes in the spatial-aware embedding to demonstrate a new spatiotemporal action retrieval scenario with composite queries. Action localization and classification experiments on four contemporary action video datasets support our proposal. Apart from state-of-the-art results in the zero-shot localization and classification settings, our spatial-aware embedding is even competitive with recent supervised action localization alternatives.

This paper aims to automatically generate a summary of an unedited video, guided by an externally provided music-track. The tempo, energy and beats in the music determine the choices and cuts in the video summarization. To solve this challenging task, we model video summarization as a quadratic assignment problem. We assign frames to the summary, using rewards based on frame interestingness, plot coherency, audio-visual match, and cut properties. Experimentally we validate our approach on the SumMe dataset. The results show that our music guided summaries are more appealing, and even outperform the current state-of-the-art summarization methods when evaluated on the F1 measure of precision and recall.

In this work, we aim to segment and detect water in videos. Water detection is beneficial for appllications such as video search, outdoor surveillance, and systems such as unmanned ground vehicles and unmanned aerial vehicles . The specific problem, however, is less discussed compared to general texture recognition. Here, we analyze several motion properties of water. First, we describe a video pre-processing step, to increase invariance against water reflections and water colours. Second, we investigate the temporal and spatial properties of water and derive corresponding local descriptors . The descriptors are used to locally classify the presence of water and a binary water detection mask is generated through spatio-temporal Markov Random Field regularization of the local classifications. Third, we introduce the Video Water Database, containing several hours of water and non-water videos, to validate our algorithm. Experimental evaluation on the Video Water Database and the DynTex database indicates the effectiveness of the proposed algorithm, outperforming multiple algorithms for dynamic texture recognition and material recognition.

This method introduces an efficient manner of learning action categories without the need of feature estimation. The approach starts from low-level values, in a similar style to the successful CNN methods. However, rather than extracting general image features, we learn to predict specific video representations from raw video data. The benefit of such an approach is that at the same computational expense it can predict 2D video representations as well as 3D ones, based on motion. The proposed model relies on discriminative Wald-boost, which we enhance to a multiclass formulation for the purpose of learning video representations. The suitability of the proposed approach as well as its time efficiency are tested on the UCF11 action recognition dataset.

This paper strives for video event detection using a representation learned from deep convolutional neural networks. Different from the leading approaches, who all learn from the 1,000 classes defined in the ImageNet Large Scale Visual Recognition Challenge, we investigate how to leverage the complete ImageNet hierarchy for pre-training deep networks. To deal with the problems of over-specific classes and classes with few images, we introduce a bottom-up and top-down approach for reorganization of the ImageNet hierarchy based on all its 21,814 classes and more than 14 million images. Experiments on the TRECVID Multimedia Event Detection 2013 and 2015 datasets show that video representations derived from the layers of a deep neural network pre-trained with our reorganized hierarchy i) improves over standard pre-training, ii) is complementary among different reorganizations, iii) maintains the benefits of fusion with other modalities, and v) leads to state-of-the-art event detection results. The reorganized hierarchies and their derived Caffe models are publicly available at http://tinyurl.com/imagenetshuffle.

To learn from visual collections, manual annotations are required. Humans however can no longer keep up with providing strong and time consuming annotations on the ever increasing wealth of visual data. As a result, approaches are required that can learn from fast and weak forms of annotations in visual data. This doctorial symposium summarizes my ongoing PhD dissertation on how to utilize weakly-supervised annotations to recognize, localize, and explain visual entities in images and videos. In this context, visual entities denote objects, scenes, and actions (in images), and actions and events (in videos). The summary is performed through four publications. For each publication, we discuss the current state-of-the-art, as well as our proposed novelties and performed experiments. The end of the summary discusses several possibilities to extend the dissertation.

The goal of this paper is event detection and recounting using a representation of concept detector scores. Different from existing work, which encodes videos by averaging concept scores over all frames, we propose to encode videos using fragments that are discriminatively learned per event. Our bag-of-fragments split a video into semantically coherent fragment proposals. From training video proposals we show how to select the most discriminative fragment for an event. An encoding of a video is in turn generated by matching and pooling these discriminative fragments to the fragment proposals of the video. The bag-of-fragments forms an effective encoding for event detection and is able to provide a precise temporally localized event recounting. Furthermore, we show how bag-of-fragments can be extended to deal with irrelevant concepts in the event recounting. Experiments on challenging web videos show that i) our modest number of fragment proposals give a high sub-event recall, ii) bag-of-fragments is complementary to global averaging and provides better event detection, iii) bag-of-fragments with concept filtering yields a desirable event recounting. We conclude that fragments matter for video event detection and recounting.