Real-world visual data rarely presents as isolated, static instances. Instead, it often evolves gradually over time through variations in pose, lighting, object state, or scene context. However, conventional classifiers are typically trained under the assumption of temporal independence, limiting their ability to capture such dynamics. We propose a simple yet effective framework that equips standard feedforward classifiers with temporal reasoning, all without modifying model architectures or introducing recurrent modules. At the heart of our approach is a novel Support-Exemplar-Query (SEQ) learning paradigm, which structures training data into temporally coherent trajectories. These trajectories enable the model to learn class-specific temporal prototypes and align prediction sequences via a differentiable soft-DTW loss. A multi-term objective further promotes semantic consistency and temporal smoothness. By interpreting input sequences as evolving feature trajectories, our method introduces a strong temporal inductive bias through loss design alone. This proves highly effective in both static and temporal tasks: it enhances performance on fine-grained and ultra-fine-grained image classification, and delivers precise, temporally consistent predictions in video anomaly detection. Despite its simplicity, our approach bridges static and temporal learning in a modular and data-efficient manner, requiring only a simple classifier on top of pre-extracted features.

Understanding human actions in videos requires more than raw pixel analysis; it relies on high-level semantic reasoning and effective integration of multimodal features. We propose a deep translational action recognition framework that enhances recognition accuracy by jointly predicting action concepts and auxiliary features from RGB video frames. At test time, hallucination streams infer missing cues, enriching feature representations without increasing computational overhead. To focus on action-relevant regions beyond raw pixels, we introduce two novel domain-specific descriptors. Object Detection Features (ODF) aggregate outputs from multiple object detectors to capture contextual cues, while Saliency Detection Features (SDF) highlight spatial and intensity patterns crucial for action recognition. Our framework seamlessly integrates these descriptors with auxiliary modalities such as optical flow, Improved Dense Trajectories, skeleton data, and audio cues. It remains compatible with state-of-the-art architectures, including I3D, AssembleNet, Video Transformer Network, FASTER, and recent models like VideoMAE V2 and InternVideo2. To handle uncertainty in auxiliary features, we incorporate aleatoric uncertainty modeling in the hallucination step and introduce a robust loss function to mitigate feature noise. Our multimodal self-supervised action recognition framework achieves state-of-the-art performance on multiple benchmarks, including Kinetics-400, Kinetics-600, and Something-Something V2, demonstrating its effectiveness in capturing fine-grained action dynamics.

Abstract Deep learning based keypoint detectors can localize specific object (or body) parts well, but still fall short of general keypoint detection. Instead, few-shot keypoint detection (FSKD) is an underexplored yet more general task of localizing either base or novel keypoints, depending on the prompted support samples. In FSKD, how to build robust keypoint representations is the key to success. To this end, we propose an FSKD approach that models relations between keypoints. As keypoints are located on objects, we exploit a class-agnostic visual prior, i.e ., the unsupervised saliency map or DINO attentiveness map to obtain the region of focus within which we perform relation learning between object patches. The class-agnostic visual prior also helps suppress the background noise largely irrelevant to keypoint locations. Then, we propose a novel Visual Prior guided Vision Transformer (VPViT). The visual prior maps are refined by a bespoke morphology learner to include relevant context of objects. The masked self-attention of VPViT takes the adapted prior map as a soft mask to constrain the self-attention to foregrounds. As robust FSKD must also deal with the low number of support samples and occlusions, based on VPViT, we further investigate i) transductive FSKD to enhance keypoint representations with unlabeled data and ii) FSKD with masking and alignment (MAA) to improve robustness. We show that our model performs well in seven public datasets, and also significantly improves the accuracy in transductive inference and under occlusions. Source codes are available at https://github.com/AlanLuSun/VPViT .

In this work, we present a novel theoretical framework for analyzing and modeling protein fitness landscapes using spectral graph theory. By representing the protein sequence space as a generalized Hamming graph and studying its spectral properties, we derive a set of powerful tools for quantifying the ruggedness, epistasis, and other key characteristics of the landscape. We prove strong approximation and sampling results, showing that the landscape can be efficiently learned and optimized from limited and noisy data. Building on this foundation, we introduce Propagational Convolutional Neural Networks (PCNNs), a new class of inductive surrogate oracle. We provide rigorous theoretical guarantees on the generalization and convergence properties of PCNNs, using techniques from the Neural Tangent Kernel framework. Extensive experiments on real-world protein engineering tasks demonstrate the superiority of PCNNs over state-of-the-art methods, achieving higher fitness and better generalization from limited data.

The challenge in open-world object detection, similarly to few- and zero-shot learning, is to generalize beyond the class distribution of the training data. In this paper, we propose a general class-agnostic objectness measure to limit bias toward labeled samples. One issue in open-world detection is that previously unseen objects are often misclassified as known categories or filtered as background by classifiers. To prevent this, we explicitly model the joint distribution of objectness and category labels using variational approximation. However, without sufficient labeled data, minimizing the KL divergence between the estimated posterior and a static normal prior fails to converge. Our theoretical analysis identifies the root cause of this failure and motivates adopting a Gaussian prior with variance dynamically adapted to the estimated posterior as a surrogate. To further reduce misclassification, we introduce an energy-based margin loss that encourages unknown objects to move toward high-density regions of the distribution, thus reducing the uncertainty of unknown detections. Our Open-World OBJectness modeling (OWOBJ) boosts novel object detection, especially in low-data regimes. OWOBJ is a flexible plugin that outperforms baselines in Open-World, Few-Shot, and zero-shot Open-Vocabulary Object Detection.

Continual learning requires models to learn tasks sequentially while maintaining a delicate balance between stability (retaining knowledge of previous tasks) and plasticity (adapting to new tasks). A key challenge is preventing interference between tasks which degrades performance when learning new tasks over previously learned tasks. Recent approaches leverage parameter-efficient fine-tuning (PEFT) which adapts pre-trained models by injecting a small number of learnable parameters. However, existing PEFT-based continual learning methods such as InfLoRA face fundamental limitations, i.e., they rely on complex optimization procedures to learn orthogonal task-specific spaces which is increasingly difficult as tasks accumulate. Thus, we propose a novel bilinear reformulation that fundamentally reimagines the task separation through fixed orthogonal bases. Our key insight is that by expanding the parameter space quadratically through two fixed bases, we can achieve "almost orthogonal" task subspaces probabilistically, eliminating the need for explicit interference elimination procedures. We provide theoretical guarantees that this approach reduces the probability of task interference from ${\mathcal{O}}\left({{{(k/d)}^2}}\right)$ to ${\mathcal{O}}\left({{{\left({k/{d^2}}\right)}^2}}\right)$, ensuring reliable task separation without complex optimization. Through extensive experiments on ImageNet-R, CIFAR-100, and DomainNet, we validate our theoretical bounds and demonstrate state-of-the-art performance with reduced parameter count. The code is available at: https://github.com/yifeiacc/BiLoRA.

In computer vision tasks, features often come from diverse representations, domains (e.g., indoor and outdoor), and modalities (e.g., text, images, and videos). Effectively fusing these features is essential for robust performance, especially with the availability of powerful pre-trained models like vision-language models. However, common fusion methods, such as concatenation, element-wise operations, and non-linear techniques, often fail to capture structural relationships, deep feature interactions, and suffer from inefficiency or misalignment of features across domains or modalities. In this paper, we shift from high-dimensional feature space to a lower-dimensional, interpretable graph space by constructing relationship graphs that encode feature relationships at different levels, e.g., clip, frame, patch, token, etc. To capture deeper interactions, we expand graphs through iterative graph relationship updates and introduce a learnable graph fusion operator to integrate these expanded relationships for more effective fusion. Our approach is relationship-centric, operates in a homogeneous space, and is mathematically principled, resembling element-wise relationship score aggregation via multilinear polynomials. We demonstrate the effectiveness of our graph-based fusion method on video anomaly detection, showing strong performance across multi-representational, multi-modal, and multi-domain feature fusion tasks.

Vision-Language Models (VLMs) such as CLIP excel at zero-shot classification due to large-scale pre-training but are vulnerable to adversarial examples. Adversarial fine-tuning robustifies zero-shot models by aligning prediction scores of individual adversaries with their clean counterparts, which typically overlooks intermediate adversarial samples along the adversarial trajectory crossing the decision boundary. Such intermediate adversaries and their vicinity produce informative representations capturing the decision boundary in detail. They can be improved by sampling adversarial candidates from simplices formed by joining two consecutive vertices on the adversarial trajectory and their clean counterpart. However, sampling simplices for adversaries is very costly. To train robust VLM, we overcome these limitations by Taylor expansion and formulating an upper-bound of alignment loss that depends on the Jacobian/Hessian obtained at clean samples. As regions between clean and intermediate adversarial samples capture a larger decision landscape, we robustify VLM by plausible adversaries from simplices by our closed-form formulation equivalent to infinite uniform sampling of the simplex. We obtain state-of-the-art robustness across 15 datasets and diverse vision-language tasks.

Mathematical diagrams have a distinctive structure. Standard feature transforms designed for natural images (e.g., CLIP) fail to process them effectively, limiting their utility in multimodal large language models (MLLMs). Current efforts to improve MLLMs have primarily focused on scaling mathematical visual instruction datasets and strengthening LLM backbones, yet fine-grained visual recognition errors remain unaddressed. Our systematic evaluation on the visual grounding capabilities of state-of-the-art MLLMs highlights that fine-grained visual understanding remains a crucial bottleneck in visual mathematical reasoning (GPT-4o exhibits a 70% grounding error rate, and correcting these errors improves reasoning accuracy by 12%). We thus propose a novel approach featuring a geometrically-grounded vision encoder and a feature router that dynamically selects between hierarchical visual feature maps. Our model accurately recognizes visual primitives and generates precise visual prompts aligned with the language model’s reasoning needs. In experiments, PRIMITIVE-Qwen2.5-7B outperforms other 7B models by 12% on MathVerse and is on par with GPT-4V on MathVista. Our findings highlight the need for better fine-grained visual integration in MLLMs. Code is available at github.com/AI4Math-ShanZhang/SVE-Math.

Contemporary Vision-Language Models (VLMs) such as CLIP offer an attractive zero-shot classification functionality facilitated by large-scale vision-language pre-training. However, they remain vulnerable to adversarial attacks, a critical security threat in realistic deployment. Adversarially robust fine-tuning provides generalizable robustness on new datasets while preserving natural performance by fine-tuning the pre-trained models. Fine-tuning robust CLIP typically relies on adversaries generated solely from the vision branch. However, this singular focus on the vision modality, coupled with static text prompts used as fixed category prototypes, limits the robustness achieved through dual-modality fine-tuning. We observe for CLIP fine-tuning that zero-shot adversarial robustness improves when we (i) stabilize the modality gap (a phenomenon where image and text features occupy different feature space regions) and (ii) lower/stabilize gradient norms. Both these steps enjoy further improvement of robustness if one fine-tunes with both visual and text adversaries. For both modalities, we leverage (i) the maximization of an effective rank of features and (ii) noise modulation of features. We show that maximizing the effective rank helps lower and stabilize the modality gap over adversaries with varying perturbation radii. The noise modulation of features, achieved by the so-called count sketching, lowers/stabilizes gradient norms. We outperform the state of the art on 15 datasets. We provide the first insights into the effects of modality gap & gradient norms in VLM fine-tuning.

We propose Graph Consistency Regularization (GCR), a novel framework that injects relational graph structures, derived from model predictions, into the learning process to promote class-aware, semantically meaningful feature representations. Functioning as a form of *self-prompting*, GCR enables the model to refine its internal structure using its own outputs. While deep networks learn rich representations, these often capture noisy inter-class similarities that contradict the model's predicted semantics. GCR addresses this issue by introducing parameter-free *Graph Consistency Layers* (GCLs) at arbitrary depths. Each GCL builds a batch-level feature similarity graph and aligns it with a global, class-aware masked prediction graph, derived by modulating softmax prediction similarities with intra-class indicators. This alignment enforces that feature-level relationships reflect class-consistent prediction behavior, acting as a *semantic regularizer* throughout the network. Unlike prior work, GCR introduces a multi-layer, cross-space graph alignment mechanism with adaptive weighting, where layer importance is learned from graph discrepancy magnitudes. This allows the model to prioritize semantically reliable layers and suppress noisy ones, enhancing feature quality without modifying the architecture or training procedure. GCR is model-agnostic, lightweight, and improves semantic structure across various networks and datasets. Experiments show that GCR promotes cleaner feature structure, stronger intra-class cohesion, and improved generalization, offering a new perspective on learning from prediction structure.

Numerous well-established studies have demonstrated the superhuman capabilities of modern Vision-Language Models (VLMs) across a wide range of tasks. However, growing is the doubt about the continuing availability of reliable high-quality labeling (supervision) from human annotators, leading to stagnation of the model's performance. To address this challenge, ``superalignment'' employs the so-called weak-to-strong generalization paradigm, where the supervision from a weak model can provide generalizable knowledge for a strong model. While effective in aligning knowledge for clean samples between the strong and weak models, the standard weak-to-strong approach typically fails to capture adversarial robustness, exposing strong VLMs to adversarial attacks. This inability to transfer adversarial robustness is because adversarial samples are normally missing in the superalignment stage. To this end, we are the first to propose the weak-to-strong (adversarial) robustness generalization method to elicit zero-shot robustness in large-scale models by an unsupervised scheme, mitigating the unreliable information source for alignment from two perspectives: alignment re-weighting and source guidance refinement. We analyze settings under which robustness generalization is possible. Extensive experiments across various vision-language benchmarks validate the effectiveness of our method in numerous scenarios, demonstrating its plug-and-play applicability to large-scale VLMs.

As foundation Vision-Language Models (VLMs) unlock fine-tuning on smaller datasets while leveraging large-scale pre-training data, machine unlearning becomes critical in addressing privacy concerns and regulatory compliance. Task vector, representing the difference between parameters of models fine-tuned with and without specific data, is a popular retraining-free unlearning strategy. However, we observe that task vectors exhibit substantial sensitivity to various fine-tuning configurations, resulting in unstable unlearning effectiveness that correlates negatively with the prediction-level variance. While aggregating multiple functions (e.g., VLM with classifier) whose parameters are represented by different task vectors reduces function variance and improves unlearning, the computational cost of obtaining numerous task vectors and aggregating functions is computationally high. Thus, in order to capture the space of task vectors induced by diverse fine-tuning strategies, we propose modeling it within the convex hull of $(Q-1)$-simplex whose vertices represent $Q$ task vectors. Although a function ensemble can be formed by sampling numerous task vectors from such a simplex, we derive a closed-form ensemble of an infinite number of functions whose parameters are uniformly sampled from the simplex, enabling efficient function-level task vector ensembling with enhanced unlearning performance. Extensive experiments and analyses across diverse datasets and scenarios demonstrate the efficacy of our method.

We introduce active generation of Pareto sets (A-GPS), a new framework for online discrete black-box multi-objective optimization (MOO). A-GPS learns a generative model of the Pareto set that supports a-posteriori conditioning on user preferences. The method employs a class probability estimator (CPE) to predict non-dominance relations and to condition the generative model toward high-performing regions of the search space. We also show that this non-dominance CPE implicitly estimates the probability of hypervolume improvement (PHVI). To incorporate subjective trade-offs, A-GPS introduces _preference direction vectors_ that encode user-specified preferences in objective space. At each iteration, the model is updated using both Pareto membership and alignment with these preference directions, producing an amortized generative model capable of sampling across the Pareto front without retraining. The result is a simple yet powerful approach that achieves high-quality Pareto set approximations, avoids explicit hypervolume computation, and flexibly captures user preferences. Empirical results on synthetic benchmarks and protein design tasks demonstrate strong sample efficiency and effective preference incorporation.

Parameter-efficient fine-tuning (PEFT) is essential for adapting large foundation models without excessive storage cost. However, current approaches such as LoRA treat each layer’s adaptation independently, overlooking correlations across layers. This independence causes the number of trainable parameters to grow linearly with model depth. We provide theoretical and empirical evidence that skip connections in transformers create smooth gradient propagation across layers. This smoothness leads to weight adaptations that concentrate most of their energy in low-frequency spectral components, especially along the layer dimension. Empirical analysis confirms this effect, showing that most of adaptation energy lies in low frequencies. Building on this insight, we propose CrossSpectra, which parameterizes all attention-weight adaptations $(Q, K, V)$ across layers as a single 3D tensor and represents them with sparse spectral coefficients ($\kappa_1, \kappa_2$). Using $\kappa_{1}$ non-zero coefficients within each layer’s frequency space and truncating to $\kappa_{2}$ frequencies across layers, CrossSpectra requires $\mathcal{O}(\kappa_{1}\kappa_{2})$ parameters instead of LoRA’s $\mathcal{O}(Lrd)$, where $L$ is the number of layers and $r$ the rank. Across natural-language and vision benchmarks, \methodname{} matches or surpasses baseline performance while using fewer parameters than LoRA, achieving only $0.36\%$ of LoRA’s parameter count when fine-tuning LLaMA-7B on instruction-following tasks. These results show that exploiting the \textbf{architectural smoothness of transformers} through spectral analysis yields major efficiency gains in PEFT.

Node classification with Graph Neural Networks (GNN) under a fixed set of labels is well studied, while Graph Few-Shot Class Incremental Learning (GFSCIL), which involves learning a GNN classifier as graph nodes and classes growing over time sporadically, has received much less attention despite its importance. We introduce inductive GFSCIL that continually learns novel classes with newly emerging nodes while maintaining performance on old classes without accessing previous data. This addresses the practical concern of transductive GFSCIL, which requires storing the entire graph with historical data. Compared to the transductive GFSCIL, the inductive setting exacerbates catastrophic forgetting due to inaccessible previous data during incremental training, in addition to the overfitting issue caused by label sparsity. Thus, we propose a novel method, called Topology-based class Augmentation and Prototype calibration (TAP). To be specific, it first performs a topology-based class augmentation method, helping replicate the setting of disjoint subgraphs with nodes of novel classes received in incremental sessions, to enhance backbone versatility. In incremental learning, given the limited number of novel class samples, we propose an iterative prototype calibration to improve the separation of class prototypes. Furthermore, as backbone fine-tuning poses the feature distribution drift, prototypes of old classes start failing over time, we propose the prototype shift method for old classes to compensate for the drift. We showcase the proposed method on four datasets.

Traditional Graph Self-Supervised Learning (GSSL) struggles to capture complex structural properties well. This limitation stems from two main factors: (1) the inadequacy of conventional Graph Neural Networks (GNNs) in representing sophisticated topological features, and (2) the focus of self-supervised learning solely on final graph representations. To address these issues, we introduce GenHopNet, a GNN framework that integrates a k-hop message-passing scheme, enhancing its ability to capture local structural information without explicit substructure extraction. We theoretically demonstrate that GenHopNet surpasses the expressiveness of the classical Weisfeiler-Lehman (WL) test for graph isomorphism. Furthermore, we propose a structural- and positional-aware GSSL framework that incorporates topological information throughout the learning process. This approach enables the learning of representations that are both sensitive to graph topology and invariant to specific structural and feature augmentations. Comprehensive experiments on graph classification datasets, including those designed to test structural sensitivity, show that our method consistently outperforms the existing approaches and maintains computational efficiency. Our work significantly advances GSSL's capability in distinguishing graphs with similar local structures but different global topologies.

Abstract Understanding human actions in videos requires more than raw pixel analysis; it relies on high-level semantic reasoning and effective integration of multimodal features. We propose a deep translational action recognition framework that enhances recognition accuracy by jointly predicting action concepts and auxiliary features from RGB video frames. At test time, hallucination streams infer missing cues, enriching feature representations without increasing computational overhead. To focus on action-relevant regions beyond raw pixels, we introduce two novel domain-specific descriptors. Object Detection Features (ODF) aggregate outputs from multiple object detectors to capture contextual cues, while Saliency Detection Features (SDF) highlight spatial and intensity patterns crucial for action recognition. Our framework seamlessly integrates these descriptors with auxiliary modalities such as optical flow, Improved Dense Trajectories, skeleton data, and audio cues. It remains compatible with state-of-the-art architectures, including I3D, AssembleNet, Video Transformer Network, FASTER, and recent models like VideoMAE V2 and InternVideo2. To handle uncertainty in auxiliary features, we incorporate aleatoric uncertainty modeling in the hallucination step and introduce a robust loss function to mitigate feature noise. Our multimodal self-supervised action recognition framework achieves state-of-the-art performance on multiple benchmarks, including Kinetics-400, Kinetics-600, and Something-Something V2, demonstrating its effectiveness in capturing fine-grained action dynamics.

Recently the prompt-based models have become popular across various language and vision tasks. Following that trend, we perform few-shot keypoint detection (FSKD) by detecting any keypoints in a query image, given the prompts formed by support images and keypoints. FSKD can be applied to detecting keypoints and poses of diverse animal species. In order to maintain flexibility of detecting varying number of keypoints, existing FSKD approaches modulate query feature map per support keypoint, then detect the corresponding keypoint from each modulated feature via a detection head. Such a separation of modulation-detection makes model heavy and slow when the number of keypoints increases. To overcome this issue, we design a novel light-weight detector which combines modulation and detection into one step, with the goal of reducing the computational cost without the drop of performance. Moreover, to bridge the large domain shift of keypoints between seen and unseen species, we further improve our model with mean feature based contrastive learning to align keypoint distributions, resulting in better keypoint representations for FSKD. Compared to the state of the art, our light-weight detector reduces the number of parameters by 50%, training/test time by 50%, and achieves 5.62% accuracy gain on 1-shot novel keypoint detection in the Animal pose dataset. Our model is also robust to the number of keypoints and saves memory when evaluating a large number of keypoints (e.g., 1000) per episode.

Adversarially robust knowledge distillation aims to com-press large-scale models into lightweight models while preserving adversarial robustness and natural performance on a given dataset. Existing methods typically align probability distributions of natural and adversarial samples between teacher and student models, but they overlook intermediate adversarial samples along the “adversarial path” formed by the multi-step gradient ascent of a sample towards the decision boundary. Such paths capture rich information about the decision boundary. In this paper, we propose a novel adversarially robust knowledge distillation approach by incorporating such adversarial paths into the alignment process. Recognizing the diverse impacts of intermediate adversarial samples (ranging from benign to noisy), we propose an adaptive weighting strategy to selectively em-phasize informative adversarial samples, thus ensuring efficient utilization of lightweight model capacity. Moreover, we propose a dual-branch mechanism exploiting two following insights: (i) complementary dynamics of adversar-ial paths obtained by targeted and untargeted adversarial learning, and (ii) inherent differences between the gradient ascent path from class$c_{i}$towards the nearest class bound-ary and the gradient descent path from a specific class$c_{j}$towards the decision region of$c_{i}(i\neq j)$. Comprehensive experiments demonstrate the effectiveness of our method on lightweight models under various settings.

Few-shot learning (FSL) facilitates a variety of computer vision tasks yet remains vulnerable to adversarial attacks. Existing adversarially robust FSL methods rely on either visual similarity learning or class concept learning. Our analysis reveals that these two learning paradigms are complementary, exhibiting distinct robustness due to their unique decision boundary types (concepts clustering by the visual similarity label vs. classification by the class labels). To bridge this gap, we propose a novel framework unifying adversarially robust similarity learning and class concept learning. Specifically, we distill parameters from both network branches into a “unified embedding model” during robust optimization and redistribute them to individual network branches periodically. To capture generalizable robustness across diverse branches, we initialize adversaries in each episode with cross-branch class-wise “global adversarial perturbations” instead of less informative random initialization. We also propose a branch robustness harmonization to modulate the optimization of similarity and class concept learners via their relative adversarial robustness. Extensive experiments demonstrate the state-of-the-art performance of our method in diverse few-shot scenarios.

Generative Adversarial Networks (GANs) significantly advanced image generation but their performance heavily depends on abundant training data. In scenarios with limited data, GANs often struggle with discriminator overfitting and unstable training. Batch Normalization (BN), despite being known for enhancing generalization and training stability, has rarely been used in the discriminator of Data-Efficient GANs. Our work addresses this gap by identifying a critical flaw in BN: the tendency for gradient explosion during the centering and scaling steps. To tackle this issue, we present CHAIN (lipsCHitz continuity constrAIned Normalization), which replaces the conventional centering step with zero-mean regularization and integrates a Lips-chitz continuity constraint in the scaling step. CHAIN further enhances GAN training by adaptively interpolating the normalized and unnormalized features, effectively avoiding discriminator overfitting. Our theoretical analyses firmly establishes CHAIN's effectiveness in reducing gradients in latent features and weights, improving stability and generalization in GAN training. Empirical evidence supports our theory. CHAIN achieves state-of-the-art results in data-limited scenarios on CIFAR-10/100, ImageNet, five low-shot and seven high-resolution few-shot image datasets.

Masked Image Modeling (MIM) is a powerful self-supervised strategy for visual pre-training without the use of labels. MIM applies random crops to input images, processes them with an encoder, and then recovers the masked inputs with a decoder, which encourages the network to capture and learn structural information about objects and scenes. The intermediate feature representations obtained from MIM are suitable for fine-tuning on downstream tasks. In this paper, we propose an Image Modeling framework based on random orthogonal projection instead of binary masking as in MIM. Our proposed Random Orthogonal Projection Image Modeling (ROPIM) reduces spatially-wise token information under guaranteed bound on the noise variance and can be considered as masking entire spatial image area under locally varying masking degrees. Since ROPIM uses a random subspace for the projection that realizes the masking step, the readily available complement of the subspace can be used during unmasking to promote recovery of removed information. In this paper, we show that using random orthogonal projection leads to superior performance compared to crop-based masking. We demonstrate state-of-the-art results on several popular benchmarks.

Contrastive learning, a form of Self-Supervised Learning (SSL), typically consists of an alignment term and a regularization term. The alignment term minimizes the distance between the embeddings of a positive pair, while the regularization term prevents trivial solutions and expresses prior beliefs about the embeddings. As a widely used regularization technique, soft decorrelation has been employed by several non-contrastive SSL methods to avoid trivial solutions. While the decorrelation term is designed to address the issue of dimensional collapse, we find that it fails to achieve this goal theoretically and experimentally. Based on such a finding, we extend the soft decorrelation regularization to minimize the distance between the covariance matrix and an identity matrix. We provide a new perspective on the geometric distance between positive definite matrices to investigate why the soft decorrelation cannot efficiently solve the dimensional collapse. Furthermore, we construct a family of loss functions utilizing the Bregman Matrix Divergence (BMD), with the soft decorrelation representing a specific instance within this family. We prove that a loss function (LogDet) in this family can solve the issue of dimensional collapse. Our novel loss functions based on BMD exhibit superior performance compared to the soft decorrelation and other baseline techniques, as demonstrated by experimental results on graph and image datasets.

Parameter-Efficient Fine-Tuning (PEFT) effectively adapts pre-trained transformers to downstream tasks. However, the optimization of tasks performance often comes at the cost of generalizability in fine-tuned models. To address this issue, we theoretically connect smaller weight gradient norms during training and larger datasets to the improvements in model generalization. Motivated by this connection, we propose reducing gradient norms for enhanced generalization and aligning fine-tuned model with the pre-trained counterpart to retain knowledge from large-scale pre-training data. Yet, naive alignment does not guarantee gradient reduction and can potentially cause gradient explosion, complicating efforts to manage gradients. To address such an issue, we propose PACE, marrying generalization of PArameter-efficient fine-tuning with Consistency rEgularization. We perturb features learned from the adapter with the multiplicative noise and ensure the fine-tuned model remains consistent for same sample under different perturbations. Theoretical analysis shows that PACE not only implicitly regularizes gradients for enhanced generalization, but also implicitly aligns the fine-tuned and pre-trained models to retain knowledge. Experimental evidence supports our theories. PACE surpasses existing PEFT methods in visual adaptation tasks (VTAB-1k, FGVC, few-shot learning, domain adaptation) showcasing its potential for resource-efficient fine-tuning. It also improves LoRA in text classification (GLUE) and mathematical reasoning (GSM-8K).

Traditional shape descriptors have been gradually replaced by convolutional neural networks due to their superior performance in feature extraction and classification. The state-of-the-art methods recognize object shapes via image reconstruction or pixel classification. However, these methods are biased toward texture information and overlook the essential shape descriptions, thus, they fail to generalize to unseen shapes. We are the first to propose a few-shot shape descriptor (FSSD) to recognize object shapes given only one or a few samples. We employ an embedding module for FSSD to extract transformation-invariant shape features. Secondly, we develop a dual attention mechanism to decompose and reconstruct the shape features via learnable shape primitives. In this way, any shape can be formed through a finite set basis, and the learned representation model is highly interpretable and extendable to unseen shapes. Thirdly, we propose a decoding module to include the supervision of shape masks and edges and align the original and reconstructed shape features, enforcing the learned features to be more shape-aware. Lastly, all the proposed modules are assembled into a few-shot shape recognition scheme. Experiments on five datasets show that our FSSD significantly improves the shape classification compared to the state-of-the-art under the few-shot setting.

Deep neural networks excel in visual recognition tasks, but their success hinges on access to balanced datasets. Yet, real-world datasets often exhibit a long-tailed distribution, compromising network efficiency and hampering generalization on unseen data. To enhance the model’s generalization in long-tailed scenarios, we present a novel feature augmentation approach termed SeMAntic tRansfer from head to Tail (SMART), which enriches the feature patterns for tail samples by transferring semantic covariance from the head classes to the tail classes along semantically correlating dimensions. This strategy boosts the model’s generalization ability by implicitly and adaptively weighting the logits, thereby widening the classification margin of tail classes. Inspired by the success of this weighting, we further incorporate a semantic-aware weighting strategy for the loss tied to tail samples. This amplifies the effect of enlarging the margin for tail classes. We are the first to provide theoretical analysis that demonstrates a large semantic diversity in tail samples can increase class margins during the training stage, leading to improved generalization. Empirical observations support our theory. Notably, with no need for extra data or learnable parameters, SMART achieves state-of-the-art results on five long-tailed benchmark datasets: CIFAR-10/100-LT, Places-LT, ImageNet-LT, and iNaturalist 2018.

Deep learning models often suffer from catastrophic forgetting when fine-tuned with samples of new classes. This issue becomes even more challenging when there is a domain shift between training and testing data. In this paper, we address the critical yet less explored Domain-Generalized Class-Incremental Learning (DGCIL) task. We propose a DGCIL approach designed to memorize old classes, adapt to new classes, and reliably classify objects from unseen domains. Specifically, our loss formulation maintains classification boundaries while suppressing domain-specific information for each class. Without storing old exemplars, we employ knowledge distillation and estimate the drift of old class prototypes as incremental training progresses. Our prototype representations are based on multivariate Normal distributions , with means and covariances continually adapted to reflect evolving model features, providing effective representations for old classes. We then sample pseudo-features for these old classes from the adapted Normal distributions using Cholesky decomposition . Unlike previous pseudo-feature sampling strategies that rely solely on average mean prototypes, our method captures richer semantic variations. Experiments on several benchmarks demonstrate the superior performance of our method compared to the state of the art.

Abstract New roads are being constructed all the time. However, the capabilities of previous deep forecasting models to generalize to new roads not seen in the training data (unseen roads) are rarely explored. In this paper, we introduce a novel setup called a spatio-temporal split to evaluate the models’ capabilities to generalize to unseen roads. In this setup, the models are trained on data from a sample of roads, but tested on roads not seen in the training data. Moreover, we also present a novel framework called Spatial Contrastive Pre-Training (SCPT) where we introduce a spatial encoder module to extract latent features from unseen roads during inference time. This spatial encoder is pre-trained using contrastive learning. During inference, the spatial encoder only requires two days of traffic data on the new roads and does not require any re-training. We also show that the output from the spatial encoder can be used effectively to infer latent node embeddings on unseen roads during inference time. The SCPT framework also incorporates a new layer, named the spatially gated addition layer, to effectively combine the latent features from the output of the spatial encoder to existing backbones. Additionally, since there is limited data on the unseen roads, we argue that it is better to decouple traffic signals to trivial-to-capture periodic signals and difficult-to-capture Markovian signals, and for the spatial encoder to only learn the Markovian signals. Finally, we empirically evaluated SCPT using the ST split setup on four real-world datasets. The results showed that adding SCPT to a backbone consistently improves forecasting performance on unseen roads. More importantly, the improvements are greater when forecasting further into the future. The codes are available on GitHub: https://github.com/cruiseresearchgroup/forecasting-on-new-roads .

Abstract Video sequences exhibit significant nuisance variations (undesired effects) of speed of actions, temporal locations, and subjects’ poses, leading to temporal-viewpoint misalignment when comparing two sets of frames or evaluating the similarity of two sequences. Thus, we propose Joint tEmporal and cAmera viewpoiNt alIgnmEnt (JEANIE) for sequence pairs. In particular, we focus on 3D skeleton sequences whose camera and subjects’ poses can be easily manipulated in 3D. We evaluate JEANIE on skeletal Few-shot Action Recognition (FSAR), where matching well temporal blocks (temporal chunks that make up a sequence) of support-query sequence pairs (by factoring out nuisance variations) is essential due to limited samples of novel classes. Given a query sequence, we create its several views by simulating several camera locations. For a support sequence, we match it with view-simulated query sequences, as in the popular Dynamic Time Warping (DTW). Specifically, each support temporal block can be matched to the query temporal block with the same or adjacent (next) temporal index, and adjacent camera views to achieve joint local temporal-viewpoint warping. JEANIE selects the smallest distance among matching paths with different temporal-viewpoint warping patterns, an advantage over DTW which only performs temporal alignment. We also propose an unsupervised FSAR akin to clustering of sequences with JEANIE as a distance measure. JEANIE achieves state-of-the-art results on NTU-60, NTU-120, Kinetics-skeleton and UWA3D Multiview Activity II on supervised and unsupervised FSAR, and their meta-learning inspired fusion.

Although augmentations (e.g., perturbation of graph edges, image crops) boost the efficiency of Contrastive Learning (CL), feature level augmentation is another plausible, complementary yet not well researched strategy. Thus, we present a novel spectral feature argumentation for contrastive learning on graphs (and images). To this end, for each data view, we estimate a low-rank approximation per feature map and subtract that approximation from the map to obtain its complement. This is achieved by the proposed herein incomplete power iteration, a non-standard power iteration regime which enjoys two valuable byproducts (under mere one or two iterations): (i) it partially balances spectrum of the feature map, and (ii) it injects the noise into rebalanced singular values of the feature map (spectral augmentation). For two views, we align these rebalanced feature maps as such an improved alignment step can focus more on less dominant singular values of matrices of both views, whereas the spectral augmentation does not affect the spectral angle alignment (singular vectors are not perturbed). We derive the analytical form for: (i) the incomplete power iteration to capture its spectrum-balancing effect, and (ii) the variance of singular values augmented implicitly by the noise. We also show that the spectral augmentation improves the generalization bound. Experiments on graph/image datasets show that our spectral feature augmentation outperforms baselines, and is complementary with other augmentation strategies and compatible with various contrastive losses.

We address the task of weakly-supervised few-shot image classification and segmentation, by leveraging a Vision Transformer (ViT) pretrained with self-supervision. Our proposed method takes token representations from the self-supervised ViT and leverages their correlations, via selfattention, to produce classification and segmentation predictions through separate task heads. Our model is able to effectively learn to perform classification and segmentation in the absence of pixel-level labels during training, using only image-level labels. To do this it uses attention maps, created from tokens generated by the self-supervised ViT backbone, as pixel-level pseudo-labels. We also explore a practical setup with “mixed” supervision, where a small number of training images contains ground-truth pixel-level labels and the remaining images have only image-level labels. For this mixed setup, we propose to improve the pseudo-labels using a pseudo-label enhancer that was trained using the available ground-truth pixel-level labels. Experiments on Pascal-5iand COCO-20idemonstrate significant performance gains in a variety of supervision settings, and in particular when little-to-no pixel-level labels are available.

Visual representation based on covariance matrix has demonstrates its efficacy for image classification by characterising the pairwise correlation of different channels in convolutional feature maps. However, pairwise correlation will become misleading once there is another channel correlating with both channels of interest, resulting in the “confounding” effect. For this case, “partial correlation” which removes the confounding effect shall be estimated instead. Nevertheless, reliably estimating partial correlation requires to solve a symmetric positive definite matrix optimisation, known as sparse inverse covariance estimation (SICE). How to incorporate this process into CNN remains an open issue. In this work, we formulate SICE as a novel structured layer of CNN. To ensure end-to-end trainability, we develop an iterative method to solve the above matrix optimisation during forward and backward propagation steps. Our work obtains a partial correlation based deep visual representation and mitigates the small sample problem often encountered by covariance matrix estimation in CNN. Computationally, our model can be effectively trained with GPU and works well with a large number of channels of advanced CNNs. Experiments show the efficacy and superior classification performance of our deep visual representation compared to covariance matrix based counterparts.

Many skeletal action recognition models use GCNs to represent the human body by 3D body joints connected body parts. GCNs aggregate one- or few-hop graph neighbourhoods, and ignore the dependency between not linked body joints. We propose to form hypergraph to model hyperedges between graph nodes (e.g., third- and fourth-order hyper-edges capture three and four nodes) which help capture higher-order motion patterns of groups of body joints. We split action sequences into temporal blocks, Higher-order Transformer (HoT) produces embeddings of each temporal block based on (i) the body joints, (ii) pairwise links of body joints and (iii) higher-order hyper-edges of skeleton body joints. We combine such HoT embeddings of hyper-edges of orders 1,…, r by a novel Multi-order Multi-mode Transformer (3Mformer) with two modules whose order can be exchanged to achieve coupled-mode attention on coupled-mode tokens based on ‘channel-temporal block’, ‘order-channel-body joint’, ‘channel-hyper-edge (any order)’ and ‘channel-only’ pairs. The first module, called Multi-order Pooling (MP), additionally learns weighted aggregation along the hyper-edge mode, whereas the second module, Temporal block Pooling (TP), aggregates along the temporal block11For brevity, we write τ temporal blocks per sequence but τ varies. mode. Our end-to-end trainable network yields state-of-the-art results compared to GCN-, transformer- and hypergraph-based counterparts.

Few-shot learning (FSL) is popular due to its ability to adapt to novel classes. Compared with inductive few-shot learning, transductive models typically perform better as they leverage all samples of the query set. The two existing classes of methods, prototype-based and graph-based, have the disadvantages of inaccurate prototype estimation and sub-optimal graph construction with kernel functions, respectively. In this paper, we propose a novel prototype-based label propagation to solve these issues. Specifically, our graph construction is based on the relation between prototypes and samples rather than between samples. As prototypes are being updated, the graph changes. We also estimate the label of each prototype instead of considering a prototype be the class centre. On mini-ImageNet, tiered-ImageNet, CIFAR-FS and CUB datasets, we show the proposed method outperforms other state-of-the-art methods in transductive FSL and semi-supervised FSL when some un-labeled data accompanies the novel few-shot task.

In instance image retrieval, considering local spatial information within an image has proven effective to boost retrieval performance, as demonstrated by local visual descriptor based geometric verification. Nevertheless, it will be highly valuable to make ordinary global image representations spatial-context-aware because global representation based image retrieval is appealing thanks to its algorithmic simplicity, low memory cost, and being friendly to sophisticated data structures. To this end, we propose a novel feature learning framework for instance image retrieval, which embeds local spatial context information into the learned global feature representations. Specifically, in parallel to the visual feature branch in a CNN backbone, we design a spatial context branch that consists of two modules called online token learning and distance encoding. For each local descriptor learned in CNN, the former module is used to indicate the types of its surrounding descriptors, while their spatial distribution information is captured by the latter module. After that, the visual feature branch and the spatial context branch are fused to produce a single global feature representation per image. As experimentally demonstrated, with the spatial-context-aware characteristic, we can well improve the performance of global representation based image retrieval while maintaining all of its appealing properties. Our code is available at https://github.com/Zy-Zhang/SpCa.

Generative Adversarial Networks (GANs) are powerful tools for image synthesis. However, they require access to vast amounts of training data, which is often costly and prohibitive. Limited data affects GANs, leading to discriminator overfitting and training instability. In this paper, we present a novel approach called NoIse-modulated Consistency rEgularization (NICE) to overcome these challenges. To this end, we introduce an adaptive multiplicative noise into the discriminator to modulate its latent features. We demonstrate the effectiveness of such a modulation in preventing discriminator overfitting by adaptively reducing the Rademacher complexity of the discriminator. However, this modulation leads to an unintended consequence of increased gradient norm, which can undermine the stability of GAN training. To mitigate this undesirable effect, we impose a constraint on the discriminator, ensuring its consistency for the same inputs under different noise modulations. The constraint effectively penalizes the first and second-order gradients of latent features, enhancing GAN stability. Experimental evidence aligns with our theoretical analysis, demonstrating the reduction of generalization error and gradient penalization of NICE. This substantiates the efficacy of NICE in reducing discriminator overfitting and improving stability of GAN training. NICE achieves state-of-the-art results on CIFAR-10, CIFAR-100, ImageNet and FFHQ datasets when trained with limited data, as well as in low-shot generation tasks.

Graph-based Collaborative Filtering (GCF) is widely used in personalized recommendation systems. However, GCF suffers from a fundamental problem where features tend to occupy the embedding space inefficiently (by spanning only a low-dimensional subspace). Such an effect is characterized in GCF by the embedding space being dominated by a few of popular items with the user embeddings highly concentrated around them. This enhances the so-called Matthew effect of the popularity bias where popular items are highly recommend whereas remaining items are ignored. In this paper, we analyze the above effect in GCF and reveal that the simplified graph convolution operation (typically used in GCF) shrinks the singular space of the feature matrix. As typical approaches (i.e., optimizing the uniformity term) fail to prevent the embedding space degradation, we propose a decorrelation-enhanced GCF objective that promotes feature diversity by leveraging the so-called principle of redundancy reduction in embeddings. However, unlike conventional methods that use the Euclidean geometry to relax hard constraints for decorrelation, we exploit non-Euclidean geometry. Such a choice helps maintain the range space of the matrix and obtain small condition number, which prevents the embedding space degradation. Our method outperforms contrastive-based GCF models on several benchmark datasets and improves the performance for unpopular items.

Traditional approaches for learning on categorical data underexploit the dependencies between columns (a.k.a. fields) in a dataset because they rely on the embedding of data points driven alone by the classification/regression loss. In contrast, we propose a novel method for learning on categorical data with the goal of exploiting dependencies between fields. Instead of modelling statistics of features globally (i.e., by the covariance matrix of features), we learn a global field dependency matrix that captures dependencies between fields and then we refine the global field dependency matrix at the instance-wise level with different weights (so-called local dependency modelling) w.r.t. each field to improve the modelling of the field dependencies. Our algorithm exploits the meta-learning paradigm, i.e., the dependency matrices are refined in the inner loop of the meta-learning algorithm without the use of labels, whereas the outer loop intertwines the updates of the embedding matrix (the matrix performing projection) and global dependency matrix in a supervised fashion (with the use of labels). Our method is simple yet it outperforms several state-of-the-art methods on six popular dataset benchmarks. Detailed ablation studies provide additional insights into our method.

We address the problem of ground-to-satellite image geo-localization, that is, estimating the camera latitude, longitude and orientation (azimuth angle) by matching a query image captured at the ground level against a large-scale database with geotagged satellite images. Our prior arts treat the above task as pure image retrieval by selecting the most similar satellite reference image matching the ground-level query image. However, such an approach often produces coarse location estimates because the geotag of the retrieved satellite image only corresponds to the image center while the ground camera can be located at any point within the image. To further consolidate our prior research finding, we present a novel geometry-aware geo-localization method. Our new method is able to achieve the fine-grained location of a query image, up to pixel size precision of the satellite image, once its coarse location and orientation have been determined. Moreover, we propose a new geometry-aware image retrieval pipeline to improve the coarse localization accuracy. Apart from a polar transform in our conference work, this new pipeline also maps satellite image pixels to the ground-level plane in the ground-view via a geometry-constrained projective transform to emphasize informative regions, such as road structures, for cross-view geo-localization. Extensive quantitative and qualitative experiments demonstrate the effectiveness of our newly proposed framework. We also significantly improve the performance of coarse localization results compared to the state-of-the-art in terms of location recalls.

We propose a Multi-level Second-order (MlSo) few-shot learning network for supervised or unsupervised few-shot image classification and few-shot action recognition. We leverage so-called power-normalized second-order base learner streams combined with features that express multiple levels of visual abstraction, and we use self-supervised discriminating mechanisms. As Second-order Pooling (SoP) is popular in image recognition, we employ its basic element-wise variant in our pipeline. The goal of multi-level feature design is to extract feature representations at different layer-wise levels of CNN, realizing several levels of visual abstraction to achieve robust few-shot learning. As SoP can handle convolutional feature maps of varying spatial sizes, we also introduce image inputs at multiple spatial scales into MlSo. To exploit the discriminative information from multi-level and multi-scale features, we develop a Feature Matching (FM) module that reweights their respective branches. We also introduce a self-supervised step, which is a discriminator of the spatial level and the scale of abstraction. Our pipeline is trained in an end-to-end manner. With a simple architecture, we demonstrate respectable results on standard datasets such as Omniglot, mini-ImageNet, tiered-ImageNet, Open MIC, fine-grained datasets such as CUB Birds, Stanford Dogs and Cars, and action recognition datasets such as HMDB51, UCF101, and mini-MIT.

Current non-rigid object keypoint detectors perform well on a chosen kind of species and body parts, and require a large amount of labelled keypoints for training. Moreover, their heatmaps, tailored to specific body parts, cannot rec-ognize novel keypoints (keypoints not labelled for training) on unseen species. We raise an interesting yet challenging question: how to detect both base (annotated for training) and novel keypoints for unseen species given a few an-notated samples? Thus, we propose a versatile Few-shot Keypoint Detection (FSKD) pipeline, which can detect a varying number of keypoints of different kinds. Our FSKD provides the uncertainty estimation of predicted keypoints. Specifically, FSKD involves main and auxiliary keypoint representation learning, similarity learning, and keypoint localization with uncertainty modeling to tackle the local-ization noise. Moreover, we model the uncertainty across groups of keypoints by multivariate Gaussian distribution to exploit implicit correlations between neighboring keypoints. We show the effectiveness of our FSKD on (i) novel keypoint detection for unseen species, (ii) few-shot Fine-Grained Vi-sual Recognition (FGVR) and (iii) Semantic Alignment (SA) downstream tasks. For FGVR, detected keypoints improve the classification accuracy. For SA, we showcase a novel thin-plate-spline warping that uses estimated keypoint un-certainty under imperfect keypoint co respondences.

In this paper11Code: https://qithub.com/MaxwellYaoNi/LCSAGAN., we improve Generative Adversarial Net-works by incorporating a manifold learning step into the discriminator. We consider locality-constrained linear and subspace-based manifolds22The coding spaces considered in this paper are loosely termed man-ifolds. In most cases they are not manifolds in the strict mathematical sense, but rather topological spaces such as varieties, or simplicial com-plexes. The word will be used only in an informal sense., and locality-constrained non-linear manifolds. In our design, the manifold learning and coding steps are intertwined with layers of the discrimina-tor, with the goal of attracting intermediate feature repre-sentations onto manifolds. We adaptively balance the dis-crepancy between feature representations and their mani-fold view, which is a trade-off between denoising on the manifold and refining the manifold. We find that locality-constrained non-linear manifolds outperform linear mani-folds due to their non-uniform density and smoothness. We also substantially outperform state-of-the-art baselines.

We design a Kernelized Few-shot Object Detector by leveraging kernelized matrices computed over multiple proposal regions, which yield expressive non-linear representations whose model complexity is learned on the fly. Our pipeline contains several modules. An Encoding Network encodes support and query images. Our Kernelized Autocorrelation unit forms the linear, polynomial and RBF kernelized representations from features extracted within support regions of support images. These features are then cross-correlated against features of a query image to obtain attention weights, and generate query proposal regions via an Attention Region Proposal Net. As the query proposal regions are many, each described by the linear, polynomial and RBF kernelized matrices, their formation is costly but that cost is reduced by our proposed Integral Region-of-Interest Aggregation unit. Finally, the Multi-head Relation Net combines all kernelized (second-order) representations with the first-order feature maps to learn support-query class relations and locations. We outperform the state of the art on novel classes by 3.8%, 5.4% and 5.7% mAP on PASCAL VOC 2007, FSOD, and COCO.

Few-shot learning (FSL) has received a lot of attention due to its remarkable ability to adapt to novel classes. Although many techniques have been proposed for FSL, they mostly focus on improving FSL backbones. Some works also focus on learning on top of the features generated by these backbones to adapt them to novel classes. We present an unsupErvised discriminAnt Subspace lEarning (EASE) that improves transductive few-shot learning performance by learning a linear projection onto a subspace built from features of the support set and the unlabeled query set in the test time. Specifically, based on the support set and the unlabeled query set, we generate the similarity matrix and the dissimilarity matrix based on the structure prior for the proposed EASE method, which is efficiently solved with SVD. We also introduce conStraIned wAsserstein MEan Shift clustEring (SIAMESE) which extends Sinkhorn K-means by incorporating labeled support samples. SIAMESE works on the features obtained from EASE to estimate class centers and query predictions. On the miniImageNet, tiered-ImageNet, CIFAR-FS, CUB and OpenMIC benchmarks, both steps significantly boost the performance in transductive FSL and semi-supervised FSL.

The main focus of existing RGB-D salient object detection models is achieving effective multi-modal fusion. Due to the limited receptive field of conventional convolutional neural networks (CNNs), CNN-based multi-modal fusion strategies fail to extensively model the correlation between the two modalities (appearance information from the RGB image and geometric information from the depth data). Given the success of transformer networks for long-range dependency modeling, we investigate multi-modal transformer networks for RGB-D salient object detection. Specifically, a transformer-based multi-modal fusion module is presented to effectively fuse appearance features and geometric features. Experimental results on six challenging benchmark RGB-D salient object detection datasets demonstrate the effectiveness of our approach.

Graph contrastive learning (GCL) improves graph representation learning, leading to SOTA on various downstream tasks. The graph augmentation step is a vital but scarcely studied step of GCL. In this paper, we show that the node embedding obtained via the graph augmentations is highly biased, somewhat limiting contrastive models from learning discriminative features for downstream tasks.Thus, instead of investigating graph augmentation in the input space, we alternatively propose to perform augmentations on the hidden features (feature augmentation). Inspired by so-called matrix sketching, we propose COSTA, a novel Covariance-preServing feaTure space Augmentation framework for GCL, which generates augmented features by maintaining a "good sketch" of original features. To highlight the superiority of feature augmentation with COSTA, we investigate a single-view setting (in addition to multi-view one) which conserves memory and computations. We show that the feature augmentation with COSTA achieves comparable/better results than graph augmentation based models.

Graph contrastive learning attracts/disperses node representations for similar/dissimilar node pairs under some notion of similarity. It may be combined with a low-dimensional embedding of nodes to preserve intrinsic and structural properties of a graph. COLES, a recent graph contrastive method combines traditional graph embedding and negative sampling into one framework. COLES in fact minimizes the trace difference between the within-class scatter matrix encapsulating the graph connectivity and the total scatter matrix encapsulating negative sampling. In this paper, we propose a more essential framework for graph embedding, called Generalized Laplacian EigeNmaps (GLEN), which learns a graph representation by maximizing the rank difference between the total scatter matrix and the within-class scatter matrix, resulting in the minimum class separation guarantee. However, the rank difference minimization is an NP-hard problem. Thus, we replace the trace difference that corresponds to the difference of nuclear norms by the difference of LogDet expressions, which we argue is a more accurate surrogate for the NP-hard rank difference than the trace difference. While enjoying a lesser computational cost, the difference of LogDet terms is lower-bounded by the Affine-invariant Riemannian metric (AIRM) and Jesen-Bregman the LogDet Divergence (JBLD), and upper-bounded by AIRM scaled by the factor of $\sqrt{m}$. We show that GLEN offers favourable accuracy/scalability compared to state-of-the-art baselines.

Even with the luxury of having abundant data, multi-label classification is widely known to be a challenging task to address. This work targets the problem of multi-label meta-learning, where a model learns to predict multiple labels within a query (e.g., an image) by just observing a few supporting examples. In doing so, we first propose a benchmark for Few-Shot Learning (FSL) with multiple labels per sample. Next, we discuss and extend several solutions specifically designed to address the conventional and single-label FSL, to work in the multi-label regime. Lastly, we introduce a neural module to estimate the label count of a given sample by exploiting the relational inference. We will show empirically the benefit of the label count module, the label propagation algorithm, and the extensions of conventional FSL methods on three challenging datasets, namely MS-COCO, iMaterialist, and Open MIC. Overall, our thorough experiments suggest that the proposed label-propagation algorithm in conjunction with the neural label count module (NLC) shall be considered as the method of choice.

Neural Architecture Search (NAS) is the game changer in designing robust neural architectures. Architectures designed by NAS outperform or compete with the best manual network designs in terms of accuracy, size, memory footprint and FLOPs. That said, previous studies focus on developing NAS algorithms for clean high quality data, a restrictive and somewhat unrealistic assumption. In this paper, focusing on the differentiable NAS algorithms, we show that vanilla NAS algorithms suffer from a performance loss if class labels are noisy. To combat this issue, we make use of the principle of information bottleneck as a regularizer. This leads us to develop a noise injecting operation that is included during the learning process, preventing the network from learning from noisy samples. Our empirical evaluations show that the noise injecting operation does not degrade the performance of the NAS algorithm if the data is indeed clean. In contrast, if the data is noisy, the architecture learned by our algorithm comfortably outperforms algorithms specifically equipped with sophisticated mechanisms to learn in the presence of label noise. In contrast to many algorithms designed to work in the presence of noisy labels, prior knowledge about the properties of the noise and its characteristics are not required for our algorithm.

Graph Neural Networks (GNNs) have achieved remarkable success by extending traditional convolution to learning on non-Euclidean data. The key to the GNNs is adopting the neural message-passing paradigm with two stages: aggregation and update. The current design of GNNs considers the topology information in the aggregation stage. However, in the updating stage, all nodes share the same updating function. The identical updating function treats each node embedding as i.i.d. random variables and thus ignores the implicit relationships between neighborhoods, which limits the capacity of the GNNs. The updating function is usually implemented with a linear transformation followed by a non-linear activation function. To make the updating function topology-aware, we inject the topological information into the non-linear activation function and propose Graph-adaptive Rectified Linear Unit (GReLU), which is a new parametric activation function incorporating the neighborhood information in a novel and efficient way. The parameters of GReLU are obtained from a hyperfunction based on both node features and the corresponding adjacent matrix. To reduce the risk of overfitting and the computational cost, we decompose the hyperfunction as two independent components for nodes and features respectively. We conduct comprehensive experiments to show that our plug-and-play GReLU method is efficient and effective given different GNN backbones and various downstream tasks.

Human actions in video sequences are characterized by the complex interplay between spatial features and their temporal dynamics. In this paper, we propose novel tensor representations for compactly capturing such higher-order relationships between visual features for the task of action recognition. We propose two tensor-based feature representations, viz. (i)sequence compatibility kernel(SCK) and (ii)dynamics compatibility kernel(DCK). SCK builds on the spatio-temporal correlations between features, whereas DCK explicitly models the action dynamics of a sequence. We also explore generalization of SCK, coined SCK$\;\oplus$, that operates on subsequences to capture the local-global interplay of correlations, which can incorporate multi-modal inputs e.g., skeleton 3D body-joints and per-frame classifier scores obtained from deep learning models trained on videos. We introduce linearization of these kernels that lead to compact and fast descriptors. We provide experiments on (i) 3D skeleton action sequences, (ii) fine-grained video sequences, and (iii) standard non-fine-grained videos. As our final representations are tensors that capture higher-order relationships of features, they relate to co-occurrences for robust fine-grained recognition (Lin, 2017), (Koniusz, 2018). We use higher-order tensors and so-called Eigenvalue Power Normalization (EPN) which have been long speculated to perform spectral detection of higher-order occurrences (Koniusz, 2013), (Koniusz, 2017), thus detecting fine-grained relationships of features rather than merely count features in action sequences. We prove that a tensor of order$r$, built from$Z_*$dimensional features, coupled with EPN indeed detects if at least one higher-order occurrence is ‘projected’ into one of its$\binom{Z_*}{r}$subspaces of dim.$r$represented by the tensor, thus forming a Tensor Power Normalization metric endowed with$\binom{Z_*}{r}$such ‘detectors’.

Power Normalizations (PN) are useful non-linear operators which tackle feature imbalances in classification problems. We study PNs in the deep learning setup via a novel PN layer pooling feature maps. Our layer combines the feature vectors and their respective spatial locations in the feature maps produced by the last convolutional layer of CNN into a positive definite matrix with second-order statistics to which PN operators are applied, forming so-called Second-order Pooling (SOP). As the main goal of this paper is to study Power Normalizations, we investigate the role and meaning of MaxExp and Gamma, two popular PN functions. To this end, we provide probabilistic interpretations of such element-wise operators and discover surrogates with well-behaved derivatives for end-to-end training. Furthermore, we look at the spectral applicability of MaxExp and Gamma by studying Spectral Power Normalizations (SPN). We show that SPN on the autocorrelation/covariance matrix and the Heat Diffusion Process (HDP) on a graph Laplacian matrix are closely related, thus sharing their properties. Such a finding leads us to the culmination of our work, a fast spectral MaxExp which is a variant of HDP for covariances/autocorrelation matrices. We evaluate our ideas on fine-grained recognition, scene recognition, and material classification, as well as in few-shot learning and graph classification.

Network embedding approaches have recently attracted considerable interest as they learn low-dimensional vector representations of nodes. Embeddings based on the matrix factorization are effective but they are usually computationally expensive due to the eigen-decomposition step. In this paper, we propose a Random RangE FInder based Network Embedding (REFINE) algorithm, which can perform embedding on one million of nodes (YouTube) within 30 seconds in a single thread. REFINE is 10x faster than ProNE, which is 10-400x faster than other methods such as LINE, DeepWalk, Node2Vec, GraRep, and Hope. Firstly, we formulate our network embedding approach as a skip-gram model, but with an orthogonal constraint, and we reformulate it into the matrix factorization problem. Instead of using randomized tSVD (truncated SVD) as other methods, we employ the Randomized Blocked QR decomposition to obtain the node representation fast. Moreover, we design a simple but efficient spectral filter for network enhancement to obtain higher-order information for node representation. Experimental results prove that REFINE is very efficient on datasets of different sizes (from thousand to million of nodes and edges) for node classification, while enjoying a good performance.

Neural networks notoriously suffer from the problem of catastrophic forgetting, the phenomenon of forgetting the past knowledge when acquiring new knowledge. Overcoming catastrophic forgetting is of significant importance to emulate the process of "incremental learning", where the model is capable of learning from sequential experience in an efficient and robust way. State-of-the-art techniques for incremental learning make use of knowledge distillation towards preventing catastrophic forgetting. Therein, one updates the network while ensuring that the network’s responses to previously seen concepts remain stable throughout updates. This in practice is done by minimizing the dissimilarity between current and previous responses of the network one way or another. Our work contributes a novel method to the arsenal of distillation techniques. In contrast to the previous state of the art, we propose to firstly construct low-dimensional manifolds for previous and current responses and minimize the dissimilarity between the responses along the geodesic connecting the manifolds. This induces a more formidable knowledge distillation with smooth properties which preserves the past knowledge more efficiently as observed by our comprehensive empirical study.1

The majority of existing few-shot learning methods describe image relations with binary labels. However, such binary relations are insufficient to teach the network complicated real-world relations, due to the lack of decision smoothness. Furthermore, current few-shot learning models capture only the similarity via relation labels, but they are not exposed to class concepts associated with objects, which is likely detrimental to the classification performance due to underutilization of the available class labels. For instance, children learn the concept of tiger from a few of actual examples as well as from comparisons of tiger to other animals. Thus, we hypothesize that both similarity and class concept learning must be occurring simultaneously. With these observations at hand, we study the fundamental problem of simplistic class modeling in current few-shot learning methods. We rethink the relations between class concepts, and propose a novel Absolute-relative Learning paradigm to fully take advantage of label information to refine the image an relation representations in both supervised and unsupervised scenarios. Our proposed paradigm improves the performance of several state-of-the-art models on publicly available datasets.

In this paper, we build on a concept of self-supervision by taking RGB frames as input to learn to predict both action concepts and auxiliary descriptors e.g., object descriptors. So-called hallucination streams are trained to predict auxiliary cues, simultaneously fed into classification layers, and then hallucinated at the testing stage to aid network. We design and hallucinate two descriptors, one leveraging four popular object detectors applied to training videos, and the other leveraging image- and video-level saliency detectors. The first descriptor encodes the detector- and Image Net-wise class prediction scores, confidence scores, and spatial locations of bounding boxes and frame indexes to capture the spatio-temporal distribution of features per video. Another descriptor encodes spatio-angular gradient distributions of saliency maps and intensity patterns. Inspired by the characteristic function of the probability distribution, we capture four statistical moments on the above intermediate descriptors. As numbers of coefficients in the mean, covariance, coskewness and cokurtotsis grow linearly, quadratically, cubically and quartically w.r.t. the dimension of feature vectors, we describe the covariance matrix by its leading n' eigenvectors (so-called subspace) and we capture skewness/kurtosis rather than costly coskewness/cokurtosis. We obtain state of the art on five popular datasets such as Charades and EPIC-Kitchens.

Graph contrastive learning attracts/disperses node representations for similar/dissimilar node pairs under some notion of similarity. It may be combined with a low-dimensional embedding of nodes to preserve intrinsic and structural properties of a graph. In this paper, we extend the celebrated Laplacian Eigenmaps with contrastive learning, and call them COntrastive Laplacian EigenmapS (COLES). Starting from a GAN-inspired contrastive formulation, we show that the Jensen-Shannon divergence underlying many contrastive graph embedding models fails under disjoint positive and negative distributions, which may naturally emerge during sampling in the contrastive setting. In contrast, we demonstrate analytically that COLES essentially minimizes a surrogate of Wasserstein distance, which is known to cope well under disjoint distributions. Moreover, we show that the loss of COLES belongs to the family of so-called block-contrastive losses, previously shown to be superior compared to pair-wise losses typically used by contrastive methods. We show on popular benchmarks/backbones that COLES offers favourable accuracy/scalability compared to DeepWalk, GCN, Graph2Gauss, DGI and GRACE baselines.

Object recognition requires a generalization capability to avoid overfitting, especially when the samples are extremely few. Generalization from limited samples, usually studied under the umbrella of meta-learning, equips learning techniques with the ability to adapt quickly in dynamical environments and proves to be an essential aspect of life long learning. In this paper, we provide a framework for few-shot learning by introducing dynamic classifiers that are constructed from few samples. A subspace method is exploited as the central block of a dynamic classifier. We will empirically show that such modelling leads to robustness against perturbations (e.g., outliers) and yields competitive results on the task of supervised and semi-supervised few-shot classification. We also develop a discriminative form which can boost the accuracy even further. Our code is available at https://github.com/chrysts/dsn_fewshot

Video-based human action recognition is currently one of the most active research areas in computer vision. Various research studies indicate that the performance of action recognition is highly dependent on the type of features being extracted and how the actions are represented. Since the release of the Kinect camera, a large number of Kinect-based human action recognition techniques have been proposed in the literature. However, there still does not exist a thorough comparison of these Kinect-based techniques under the grouping of feature types, such as handcrafted versus deep learning features and depth-based versus skeleton-based features. In this paper, we analyze and compare 10 recent Kinect-based algorithms for both cross-subject action recognition and cross-view action recognition using six benchmark datasets. In addition, we have implemented and improved some of these techniques and included their variants in the comparison. Our experiments show that the majority of methods perform better on cross-subject action recognition than cross-view action recognition, that the skeleton-based features are more robust for cross-view recognition than the depth-based features, and that the deep learning features are suitable for large datasets.

Despite deep end-to-end learning methods have shown their superiority in removing non-uniform motion blur, there still exist major challenges with the current multi-scale and scale-recurrent models: 1) Deconvolution/upsampling operations in the coarse-to-fine scheme result in expensive runtime; 2) Simply increasing the model depth with finer-scale levels cannot improve the quality of deblurring. To tackle the above problems, we present a deep {hierarchical multi-patch network} inspired by Spatial Pyramid Matching to deal with blurry images via a fine-to-coarse hierarchical representation. To deal with the performance saturation w.r.t. depth, we propose a stacked version of our multi-patch model. Our proposed basic multi-patch model achieves the state-of-the-art performance on the GoPro dataset while enjoying a 40$\times$ faster runtime compared to current multi-scale methods. With 30ms to process an image at 1280$\times$720 resolution, it is the first real-time deep motion deblurring model for 720p images at 30fps. For stacked networks, significant improvements (over 1.2dB) are achieved on the GoPro dataset by increasing the network depth. Moreover, by varying the depth of the stacked model, one can adapt the performance and runtime of the same network for different application scenarios.

Learning new concepts from a few of samples is a standard challenge in computer vision. The main directions to improve the learning ability of few-shot training models include (i) a robust similarity learning and (ii) generating or hallucinating additional data from the limited existing samples. In this paper, we follow the latter direction and present a novel data hallucination model. Currently, most datapoint generators contain a specialized network (i.e., GAN) tasked with hallucinating new datapoints, thus requiring large numbers of annotated data for their training in the first place. In this paper, we propose a novel less-costly hallucination method for few-shot learning which utilizes saliency maps. To this end, we employ a saliency network to obtain the foregrounds and backgrounds of available image samples and feed the resulting maps into a two-stream network to hallucinate datapoints directly in the feature space from viable foreground-background combinations. To the best of our knowledge, we are the first to leverage saliency maps for such a task and we demonstrate their usefulness in hallucinating additional datapoints for few-shot learning. Our proposed network achieves the state of the art on publicly available datasets.

The prediction of flight delays plays a significantly important role for airlines and travellers because flight delays cause not only tremendous economic loss but also potential security risks. In this work, we aim to integrate multiple data sources to predict the departure delay of a scheduled flight. Different from previous work, we are the first group, to our best knowledge, to take advantage of airport situational awareness map, which is defined as airport traffic complexity (ATC), and combine the proposed ATC factors with weather conditions and light information. Features engineering methods and most state-of-the-art machine learning algorithms are applied to a large real-world data sources. We reveal a couple of factors at the airport which has a significant impact on flight departure delay time. The prediction results show that the proposed factors are the main reasons behind the flight delays. Using our proposed framework, an improvement in accuracy for flight departure delay prediction is obtained.

In this paper, we revive the use of old-fashioned handcrafted video representations for action recognition and put new life into these techniques via a CNN-based hallucination step. Despite of the use of RGB and optical flow frames, the I3D model (amongst others) thrives on combining its output with the Improved Dense Trajectory (IDT) and extracted with its low-level video descriptors encoded via Bag-of-Words (BoW) and Fisher Vectors (FV). Such a fusion of CNNs and handcrafted representations is time-consuming due to pre-processing, descriptor extraction, encoding and tuning parameters. Thus, we propose an end-to-end trainable network with streams which learn the IDT-based BoW/FV representations at the training stage and are simple to integrate with the I3D model. Specifically, each stream takes I3D feature maps ahead of the last 1D conv. layer and learns to `translate' these maps to BoW/FV representations. Thus, our model can hallucinate and use such synthesized BoW/FV representations at the testing stage. We show that even features of the entire I3D optical flow stream can be hallucinated thus simplifying the pipeline. Our model saves 20-55h of computations and yields state-of-the-art results on four publicly available datasets.

Recovering a photorealistic face from an artistic portrait is a challenging task since crucial facial details are often distorted or completely lost in artistic compositions. To handle this loss, we propose an Attribute-guided Face Recovery from Portraits (AFRP) that utilizes a Face Recovery Network (FRN) and a Discriminative Network (DN). FRN consists of an autoencoder with residual block-embedded skip-connections and incorporates facial attribute vectors into the feature maps of input portraits at the bottleneck of the autoencoder. DN has multiple convolutional and fully-connected layers, and its role is to enforce FRN to generate authentic face images with corresponding facial attributes dictated by the input attribute vectors. For the preservation of identities, we impose the recovered and ground-truth faces to share similar visual features. Specifically, DN determines whether the recovered image looks like a real face and checks if the facial attributes extracted from the recovered image are consistent with given attributes. Our method can recover photorealistic identity-preserving faces with desired attributes from unseen stylized portraits, artistic paintings, and hand-drawn sketches. On large-scale synthesized and sketch datasets, we demonstrate that our face recovery method achieves state-of-the-art results.

Second-and higher-order statistics of data points have played an important role in advancing the state of the art on several computer vision problems such as the fine-grained image and scene recognition. However, these statistics need to be passed via an appropriate pooling scheme to obtain the best performance. Power Normalizations are non-linear activation units which enjoy probability-inspired derivations and can be applied in CNNs. In this paper, we propose a similarity learning network leveraging second-order information and Power Normalizations. To this end, we propose several formulations capturing second-order statistics and derive a sigmoid-like Power Normalizing function to demonstrate its interpretability. Our model is trained end-to-end to learn the similarity between the support set and query images for the problem of one-and few-shot learning. The evaluations on Omniglot, miniImagenet and Open MIC datasets demonstrate that this network obtains state-of-the-art results on several few-shot learning protocols.

Power Normalizations (PN) are very useful non-linear operators in the context of Bag-of-Words data representations as they tackle problems such as feature imbalance. In this paper, we reconsider these operators in the deep learning setup by introducing a novel layer that implements PN for non-linear pooling of feature maps. Specifically, by using a kernel formulation, our layer combines the feature vectors and their respective spatial locations in the feature maps produced by the last convolutional layer of CNN. Linearization of such a kernel results in a positive definite matrix capturing the second-order statistics of the feature vectors, to which PN operators are applied. We study two types of PN functions, namely (i) MaxExp and (ii) Gamma, addressing their role and meaning in the context of nonlinear pooling. We also provide a probabilistic interpretation of these operators and derive their surrogates with well-behaved gradients for end-to-end CNN learning. We apply our theory to practice by implementing the PN layer on a ResNet-50 model and showcase experiments on four benchmarks for fine-grained recognition, scene recognition, and material classification. Our results demonstrate state-of-the-part performance across all these tasks.

In this paper, we address an open problem of zero-shot learning. Its principle is based on learning a mapping that associates feature vectors extracted from i.e. images and attribute vectors that describe objects and/or scenes of interest. In turns, this allows classifying unseen object classes and/or scenes by matching feature vectors via mapping to a newly defined attribute vector describing a new class. Due to importance of such a learning task, there exist many methods that learn semantic, probabilistic, linear or piece-wise linear mappings. In contrast, we apply well-established kernel methods to learn a non-linear mapping between the feature and attribute spaces. We propose an easy learning objective inspired by the Linear Discriminant Analysis, Kernel-Target Alignment and Kernel Polarization methods [12, 8, 4] that promotes incoherence. We evaluate the performance of our algorithm on the Polynomial as well as shift-invariant Gaussian and Cauchy kernels. Despite simplicity of our approach, we obtain state-of-the-art results on several zero-shot learning datasets and benchmarks including a recent AWA2 dataset [45].

Recovering the latent photorealistic faces from their artistic portraits aids human perception and facial analysis. However, a recovery process that can preserve identity is challenging because the fine details of real faces can be distorted or lost in stylized images. In this paper, we present a new Identity-preserving Face Recovery from Portraits (IFRP) to recover latent photorealistic faces from unaligned stylized portraits. Our IFRP method consists of two components: Style Removal Network (SRN) and Discriminative Network (DN). The SRN is designed to transfer feature maps of stylized images to the feature maps of the corresponding photorealistic faces. By embedding spatial transformer networks into the SRN, our method can compensate for misalignments of stylized faces automatically and output aligned realistic face images. The role of the DN is to enforce recovered faces to be similar to authentic faces. To ensure the identity preservation, we promote the recovered and ground-truth faces to share similar visual features via a distance measure which compares features of recovered and ground-truth faces extracted from a pre-trained VGG network. We evaluate our method on a large-scale synthesized dataset of real and stylized face pairs and attain state of the art results. In addition, our method can recover photorealistic faces from previously unseen stylized portraits, original paintings and human-drawn sketches.

In this paper, we propose an approach to the domain adaptation, dubbed Second-or Higher-order Transfer of Knowledge (So-HoT), based on the mixture of alignments of second-or higher-order scatter statistics between the source and target domains. The human ability to learn from few labeled samples is a recurring motivation in the literature for domain adaptation. Towards this end, we investigate the supervised target scenario for which few labeled target training samples per category exist. Specifically, we utilize two CNN streams: the source and target networks fused at the classifier level. Features from the fully connected layers fc7 of each network are used to compute second-or even higher-order scatter tensors, one per network stream per class. As the source and target distributions are somewhat different despite being related, we align the scatters of the two network streams of the same class (within-class scatters) to a desired degree with our bespoke loss while maintaining good separation of the between-class scatters. We train the entire network in end-to-end fashion. We provide evaluations on the standard Office benchmark (visual domains) and RGB-D combined with Caltech256 (depth-to-rgb transfer). We attain state-of-the-art results.

Most successful deep learning algorithms for action recognition extend models designed for image-based tasks such as object recognition to video. Such extensions are typically trained for actions on single video frames or very short clips, and then their predictions from sliding-windows over the video sequence are pooled for recognizing the action at the sequence level. Usually this pooling step uses the first-order statistics of frame-level action predictions. In this paper, we explore the advantages of using higherorder correlations, specifically, we introduce Higher-order Kernel (HOK) descriptors generated from the late fusion of CNN classifier scores from all the frames in a sequence. To generate these descriptors, we use the idea of kernel linearization. Specifically, a similarity kernel matrix, which captures the temporal evolution of deep classifier scores, is first linearized into kernel feature maps. The HOK descriptors are then generated from the higher-order cooccurrences of these feature maps, and are then used as input to a video-level classifier. We provide experiments on two fine-grained action recognition datasets, and show that our scheme leads to state-of-the-art results.

In object recognition, the Bag-of-Words model assumes: i) extraction of local descriptors from images, ii) embedding the descriptors by a coder to a given visual vocabulary space which results in mid-level features, iii) extracting statistics from mid-level features with a pooling operator that aggregates occurrences of visual words in images into signatures, which we refer to as First-order Occurrence Pooling. This paper investigates higher-order pooling that aggregates over co-occurrences of visual words. We derive Bag-of-Words with Higher-order Occurrence Pooling based on linearisation of Minor Polynomial Kernel, and extend this model to work with various pooling operators. This approach is then effectively used for fusion of various descriptor types. Moreover, we introduce Higher-order Occurrence Pooling performed directly on local image descriptors as well as a novel pooling operator that reduces the correlation in the image signatures. Finally, First-, Second-, and Third-order Occurrence Pooling are evaluated given various coders and pooling operators on several widely used benchmarks. The proposed methods are compared to other approaches such as Fisher Vector Encoding and demonstrate improved results.

Super-symmetric tensors - a higher-order extension of scatter matrices - are becoming increasingly popular in machine learning and computer vision for modeling data statistics, co-occurrences, or even as visual descriptors. They were shown recently to outperform second-order approaches, however, the size of these tensors are exponential in the data dimensionality, which is a significant concern. In this paper, we study third-order supersymmetric tensor descriptors in the context of dictionary learning and sparse coding. For this purpose, we propose a novel non-linear third-order texture descriptor. Our goal is to approximate these tensors as sparse conic combinations of atoms from a learned dictionary. Apart from the significant benefits to tensor compression that this framework offers, our experiments demonstrate that the sparse coefficients produced by this scheme lead to better aggregation of high-dimensional data and showcase superior performance on two common computer vision tasks compared to the state of the art.

We investigate the problem of visual tracking of multiple human speakers in an office environment. In particular, we propose novel solutions to the following challenges: (1) robust and computationally efficient modeling and classification of the changing appearance of the speakers in a variety of different lighting conditions and camera resolutions; (2) dealing with full or partial occlusions when multiple speakers cross or come into very close proximity; (3) automatic initialization of the trackers, or re-initialization when the trackers have lost lock caused by e.g. the limited camera views. First, we develop new algorithms for appearance modeling of the moving speakers based on dictionary learning (DL), using an off-line training process. In the tracking phase, the histograms (coding coefficients) of the image patches derived from the learned dictionaries are used to generate the likelihood functions based on Support Vector Machine (SVM) classification. This likelihood function is then used in the measurement step of the classical particle filtering (PF) algorithm. To improve the computational efficiency of generating the histograms, a soft voting technique based on approximate Locality-constrained Soft Assignment (LcSA) is proposed to reduce the number of dictionary atoms (codewords) used for histogram encoding. Second, an adaptive identity model is proposed to track multiple speakers whilst dealing with occlusions. This model is updated online using Maximum a Posteriori (MAP) adaptation, where we control the adaptation rate using the spatial relationship between the subjects. Third, to enable automatic initialization of the visual trackers, we exploit audio information, the Direction of Arrival (DOA) angle, derived from microphone array recordings. Such information provides, a priori, the number of speakers and constrains the search space for the speaker's faces. The proposed system is tested on a number of sequences from three publicly available and challenging data corpora (AV16.3, EPFL pedestrian data set and CLEAR) with up to five moving subjects.

Visual concept detection and action recognition are one of the most important tasks in content-based multimedia information retrieval (CBMIR) technology. It aims at annotating images using a vocabulary defined by a set of concepts of interest including scenes types (mountains, snow, etc.) or human actions (phoning, playing instrument). This paper describes our system in the ImageCLEF@ICPR10, Pascal VOC 08 Visual Concept Detection and Pascal VOC 10 Action Recognition Challenges. The proposed system ranked first in these large-scale tasks when evaluated independently by the organizers. The proposed system involves state-of-the-art local descriptor computation, vector quantization via clustering, structured scene or object representation via localized histograms of vector codes, similarity measure for kernel construction and classifier learning. The main novelty is the classifier-level and kernel-level fusion using Kernel Discriminant Analysis and Spectral Regression (SR-KDA) with RBF Chi-Squared kernels obtained from various image descriptors. The distinctiveness of the proposed method is also assessed experimentally using a video benchmark: the Mediamill Challenge along with benchmarks from ImageCLEF@ICPR10, Pascal VOC 10 and Pascal VOC 08. From the experimental results, it can be derived that the presented system consistently yields significant performance gains when compared with the state-of-the art methods. The other strong point is the introduction of SR-KDA in the classification stage where the time complexity scales linearly with respect to the number of concepts and the main computational complexity is independent of the number of categories.

Spatial Pyramid Match lies at a heart of modern object category recognition systems. Once image descriptors are expressed as histograms of visual words, they are further deployed across spatial pyramid with coarse-to-fine spatial location grids. However, such representation results in extreme histogram vectors of 200K or more elements increasing computational and memory requirements. This paper investigates alternative ways of introducing spatial information during formation of histograms. Specifically, we propose to apply spatial location information at a descriptor level and refer to it as Spatial Coordinate Coding. Alternatively, x, y, radius, or angle is used to perform semi-coding. This is achieved by adding one of the spatial components at the descriptor level whilst applying Pyramid Match to another. Lastly, we demonstrate that Pyramid Match can be applied robustly to other measurements: Dominant Angle and Colour. We demonstrate state-of-the art results on two datasets with means of Soft Assignment and Sparse Coding.

Visual Word Uncertainty also referred to as Soft Assignment is a well established technique for representing images as histograms by flexible assignment of image descriptors to a visual vocabulary. Recently, an attention of the community dealing with the object category recognition has been drawn to Linear Coordinate Coding methods. In this work, we focus on Soft Assignment as it yields good results amidst competitive methods. We show that one can take two views on Soft Assignment: an approach derived from Gaussian Mixture Model or special case of Linear Coordinate Coding. The latter view helps us propose how to optimise smoothing factor of Soft Assignment in a way that minimises descriptor reconstruction error and maximises classification performance. In turns, this renders tedious cross-validation towards establishing this parameter unnecessary and yields it a handy technique. We demonstrate state-of-the-art performance of such optimised assignment on two image datasets and several types of descriptors.

This paper investigates segmentation-based image descriptors for object category recognition. In contrast to commonly used interest points the proposed descriptors are extracted from pairs of adjacent regions given by a segmentation method. In this way we exploit semi-local structural information from the image. We propose to use the segments as spatial bins for descriptors of various image statistics based on gradient, colour and region shape. Proposed descriptors are validated on standard recognition benchmarks. Results show they outperform state-of-the-art reference descriptors with 5.6x less data and achieve comparable results to them with 8.6x less data. The proposed descriptors are complementary to SIFT and achieve state-of-the-art results when combined together within a kernel based classifier.

This paper investigates segmentation based interest points for matching and recognition.We propose two simple methods for extracting features from the segmentation maps, which focus on the boundaries and centres of the gravity of the segments.In addition, this can be considered a novel approach for evaluating unsupervised image segmentation algorithms.Former evaluations aim at estimating segmentation quality by how well resulting segments adhere to the contours separating ground-truth foregrounds from backgrounds and therefore explicitly focus on particular objects of interest.In contrast, we propose to measure the robustness of segmentations by the repeatability of features extracted from segments on images related by various geometric and photometric transformations.Further, our evaluation provides a new insight into suitability of the segmentation methods for generating local features for image retrieval or recognition.Several segmentation methods are evaluated and compared to state-of-the art interest point detectors using the repeatability criteria as well as standard matching and recognition benchmarks.