Vittorio Murino

Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota et al.

Over the last few years, Unsupervised Domain Adaptation (UDA) techniques have acquired remarkable importance and popularity in computer vision. However, when compared to the extensive literature available for images, the field of videos is still relatively unexplored. On the other hand, the performance of a model in action recognition is heavily affected by domain shift. In this paper, we propose a simple and novel UDA approach for video action recognition. Our approach leverages recent advances on spatio-temporal transformers to build a robust source model that better generalises to the target domain. Furthermore, our architecture learns domain invariant features thanks to the introduction of a novel alignment loss term derived from the Information Bottleneck principle. We report results on two video action recognition benchmarks for UDA, showing state-of-the-art performance on HMDB$\leftrightarrow$UCF, as well as on Kinetics$\rightarrow$NEC-Drone, which is more challenging. This demonstrates the effectiveness of our method in handling different levels of domain shift. The source code is available at https://github.com/vturrisi/UDAVT.

Andrea Toaiari, Vittorio Murino, Marco Cristani et al.

Gaze Target Detection (GTD), i.e., determining where a person is looking within a scene from an external viewpoint, is a challenging task, particularly in 3D space. Existing approaches heavily rely on analyzing the person's appearance, primarily focusing on their face to predict the gaze target. This paper presents a novel approach to tackle this problem by utilizing the person's upper-body pose and available depth maps to extract a 3D gaze direction and employing a multi-stage or an end-to-end pipeline to predict the gazed target. When predicted accurately, the human body pose can provide valuable information about the head pose, which is a good approximation of the gaze direction, as well as the position of the arms and hands, which are linked to the activity the person is performing and the objects they are likely focusing on. Consequently, in addition to performing gaze estimation in 3D, we are also able to perform GTD simultaneously. We demonstrate state-of-the-art results on the most comprehensive publicly accessible 3D gaze target detection dataset without requiring images of the person's face, thus promoting privacy preservation in various application contexts. The code is available at https://github.com/intelligolabs/privacy-gtd-3D.

Julio Ivan Davila Carrazco, Pietro Morerio, Alessio Del Bue et al.

This paper presents a classification framework based on learnable data augmentation to tackle the One-Shot Unsupervised Domain Adaptation (OS-UDA) problem. OS-UDA is the most challenging setting in Domain Adaptation, as only one single unlabeled target sample is assumed to be available for model adaptation. Driven by such single sample, our method LearnAug-UDA learns how to augment source data, making it perceptually similar to the target. As a result, a classifier trained on such augmented data will generalize well for the target domain. To achieve this, we designed an encoder-decoder architecture that exploits a perceptual loss and style transfer strategies to augment the source data. Our method achieves state-of-the-art performance on two well-known Domain Adaptation benchmarks, DomainNet and VisDA. The project code is available at https://github.com/IIT-PAVIS/LearnAug-UDA

Sanket Thakur, Cigdem Beyan, Pietro Morerio et al.

Objects are crucial for understanding human-object interactions. By identifying the relevant objects, one can also predict potential future interactions or actions that may occur with these objects. In this paper, we study the problem of Short-Term Object interaction anticipation (STA) and propose NAOGAT (Next-Active-Object Guided Anticipation Transformer), a multi-modal end-to-end transformer network, that attends to objects in observed frames in order to anticipate the next-active-object (NAO) and, eventually, to guide the model to predict context-aware future actions. The task is challenging since it requires anticipating future action along with the object with which the action occurs and the time after which the interaction will begin, a.k.a. the time to contact (TTC). Compared to existing video modeling architectures for action anticipation, NAOGAT captures the relationship between objects and the global scene context in order to predict detections for the next active object and anticipate relevant future actions given these detections, leveraging the objects' dynamics to improve accuracy. One of the key strengths of our approach, in fact, is its ability to exploit the motion dynamics of objects within a given clip, which is often ignored by other models, and separately decoding the object-centric and motion-centric information. Through our experiments, we show that our model outperforms existing methods on two separate datasets, Ego4D and EpicKitchens-100 ("Unseen Set"), as measured by several additional metrics, such as time to contact, and next-active-object localization. The code will be available upon acceptance.

Vito Paolo Pastore, Massimiliano Ciranni, Davide Marinelli et al.

It is widely recognized that deep neural networks are sensitive to bias in the data. This means that during training these models are likely to learn spurious correlations between data and labels, resulting in limited generalization abilities and low performance. In this context, model debiasing approaches can be devised aiming at reducing the model's dependency on such unwanted correlations, either leveraging the knowledge of bias information or not. In this work, we focus on the latter and more realistic scenario, showing the importance of accurately predicting the bias-conflicting and bias-aligned samples to obtain compelling performance in bias mitigation. On this ground, we propose to conceive the problem of model bias from an out-of-distribution perspective, introducing a new bias identification method based on anomaly detection. We claim that when data is mostly biased, bias-conflicting samples can be regarded as outliers with respect to the bias-aligned distribution in the feature space of a biased model, thus allowing for precisely detecting them with an anomaly detection method. Coupling the proposed bias identification approach with bias-conflicting data upsampling and augmentation in a two-step strategy, we reach state-of-the-art performance on synthetic and real benchmark datasets. Ultimately, our proposed approach shows that the data bias issue does not necessarily require complex debiasing methods, given that an accurate bias identification procedure is defined. Source code is available at https://github.com/Malga-Vision/MoDAD

Sanket Thakur, Cigdem Beyan, Pietro Morerio et al.

This paper addresses the problem of anticipating the next-active-object location in the future, for a given egocentric video clip where the contact might happen, before any action takes place. The problem is considerably hard, as we aim at estimating the position of such objects in a scenario where the observed clip and the action segment are separated by the so-called ``time to contact'' (TTC) segment. Many methods have been proposed to anticipate the action of a person based on previous hand movements and interactions with the surroundings. However, there have been no attempts to investigate the next possible interactable object, and its future location with respect to the first-person's motion and the field-of-view drift during the TTC window. We define this as the task of Anticipating the Next ACTive Object (ANACTO). To this end, we propose a transformer-based self-attention framework to identify and locate the next-active-object in an egocentric clip. We benchmark our method on three datasets: EpicKitchens-100, EGTEA+ and Ego4D. We also provide annotations for the first two datasets. Our approach performs best compared to relevant baseline methods. We also conduct ablation studies to understand the effectiveness of the proposed and baseline methods on varying conditions. Code and ANACTO task annotations will be made available upon paper acceptance.

Fanqi Yu, Matteo Tiezzi, Tommaso Apicella et al.

We introduce a lifelong imitation learning framework that enables continual policy refinement across sequential tasks under realistic memory and data constraints. Our approach departs from conventional experience replay by operating entirely in a multimodal latent space, where compact representations of visual, linguistic, and robot's state information are stored and reused to support future learning. To further stabilize adaptation, we introduce an incremental feature adjustment mechanism that regularizes the evolution of task embeddings through an angular margin constraint, preserving inter-task distinctiveness. Our method establishes a new state of the art in the LIBERO benchmarks, achieving 10-17 point gains in AUC and up to 65% less forgetting compared to previous leading methods. Ablation studies confirm the effectiveness of each component, showing consistent gains over alternative strategies. The code is available at: https://github.com/yfqi/lifelong_mlr_ifa.

Waqar Ahmed, Pietro Morerio, Vittorio Murino

Existing Source-free Unsupervised Domain Adaptation (SUDA) approaches inherently exhibit catastrophic forgetting. Typically, models trained on a labeled source domain and adapted to unlabeled target data improve performance on the target while dropping performance on the source, which is not available during adaptation. In this study, our goal is to cope with the challenging problem of SUDA in a continual learning setting, i.e., adapting to the target(s) with varying distributional shifts while maintaining performance on the source. The proposed framework consists of two main stages: i) a SUDA model yielding cleaner target labels -- favoring good performance on target, and ii) a novel method for synthesizing class-conditioned source-style images by leveraging only the source model and pseudo-labeled target data as a prior. An extensive pool of experiments on major benchmarks, e.g., PACS, Visda-C, and DomainNet demonstrates that the proposed Continual SUDA (C-SUDA) framework enables preserving satisfactory performance on the source domain without exploiting the source data at all.

Massimiliano Ciranni, Luca Molinaro, Carlo Alberto Barbano et al.

In the last few years, due to the broad applicability of deep learning to downstream tasks and end-to-end training capabilities, increasingly more concerns about potential biases to specific, non-representative patterns have been raised. Many works focusing on unsupervised debiasing usually leverage the tendency of deep models to learn ``easier'' samples, for example by clustering the latent space to obtain bias pseudo-labels. However, the interpretation of such pseudo-labels is not trivial, especially for a non-expert end user, as it does not provide semantic information about the bias features. To address this issue, we introduce ``Say My Name'' (SaMyNa), the first tool to identify biases within deep models semantically. Unlike existing methods, our approach focuses on biases learned by the model. Our text-based pipeline enhances explainability and supports debiasing efforts: applicable during either training or post-hoc validation, our method can disentangle task-related information and proposes itself as a tool to analyze biases. Evaluation on traditional benchmarks demonstrates its effectiveness in detecting biases and even disclaiming them, showcasing its broad applicability for model diagnosis.

Pietro Morerio, Ruggero Ragonesi, Vittorio Murino

Deep Neural Networks are well known for efficiently fitting training data, yet experiencing poor generalization capabilities whenever some kind of bias dominates over the actual task labels, resulting in models learning "shortcuts". In essence, such models are often prone to learn spurious correlations between data and labels. In this work, we tackle the problem of learning from biased data in the very realistic unsupervised scenario, i.e., when the bias is unknown. This is a much harder task as compared to the supervised case, where auxiliary, bias-related annotations, can be exploited in the learning process. This paper proposes a novel 2-stage learning pipeline featuring a data augmentation strategy able to regularize the training. First, biased/unbiased samples are identified by training over-biased models. Second, such subdivision (typically noisy) is exploited within a data augmentation framework, properly combining the original samples while learning mixing parameters, which has a regularization effect. Experiments on synthetic and realistic biased datasets show state-of-the-art classification accuracy, outperforming competing methods, ultimately proving robust performance on both biased and unbiased examples. Notably, being our training method totally agnostic to the level of bias, it also positively affects performance for any, even apparently unbiased, dataset, thus improving the model generalization regardless of the level of bias (or its absence) in the data.

Francesco Olivato, Cigdem Beyan, Vittorio Murino

In this work, we study Source-Free Unsupervised Domain Adaptation under corruption-induced domain shifts, where performance degradation is caused by natural image corruptions that go beyond additive noise, including blur, weather effects, and digital artifacts. We propose a diffusion-based, input-level adaptation framework that operates entirely at test time and keeps all source-trained models frozen, explicitly targeting robustness to corrupted target inputs. Our method leverages a source-trained diffusion model as a generative prior and introduces a discriminator-guided adaptive diffusion strategy that dynamically controls the amount of perturbation applied to each test sample. Rather than relying on a fixed diffusion depth, the discriminator determines, on a per-image basis, when sufficient forward diffusion has been applied to suppress corruption-specific artifacts, with each corruption type effectively defining a distinct target domain. This adaptive stopping mechanism applies only the necessary amount of noise to remove domainspecific corruption while preserving class-discriminative structure. The reverse diffusion process then reconstructs a source-aligned image, optionally stabilized through structural guidance, which is classified using a frozen source-trained classifier. We evaluate the proposed approach across a broad spectrum of corruption-induced target domains, covering 15 diverse corruption types, and demonstrate more balanced robustness with competitive or improved performance across non-noise corruptions. Additional analyses reveal how the adaptive diffusion schedule responds to different corruption characteristics, highlighting the practicality, generality, and robustness of the proposed framework. The code is publicly available at https://github.com/fmolivato/dgadiffusion/.

Francesco Dibitonto, Cigdem Beyan, Vittorio Murino

Recent advances in representation learning have shown that hyperbolic geometry can offer a more expressive alternative to the Euclidean embeddings used in CLIP models, capturing hierarchical structures and leading to better-organized representations. However, current hyperbolic CLIP variants are trained entirely from scratch, which is computationally expensive and resource-intensive. In this work, we propose HAC (Hyperbolic Adaptation of CLIP), a parameter-efficient framework that enables pretrained CLIP models to transition into hyperbolic space via lightweight fine-tuning. We apply HAC to Visual Question Answering (VQA), where models must interpret visual elements and align them with textual queries. Notably, HAC's training is performed on a dataset with no overlap with any VQA benchmark, resulting in a strict zero-shot evaluation paradigm that underscores HAC's task-agnostic adaptability. We evaluate HAC across a diverse suite of VQA benchmarks spanning General, Reasoning, and OCR categories. Both HAC-S (small) and HAC-B (medium) consistently surpass Euclidean baselines and prior hyperbolic approaches, with HAC-B delivering up to a +1.9 point average improvement over CLIP-B on reasoning-intensive tasks. Our code is available at https://github.com/fdibiton/HAC

Mahshid Shiri, Cigdem Beyan, Vittorio Murino

An innovative few-shot anomaly detection approach is presented, leveraging the pre-trained CLIP model for medical data, and adapting it for both image-level anomaly classification (AC) and pixel-level anomaly segmentation (AS). A dual-branch design is proposed to separately capture normal and abnormal features through learnable adapters in the CLIP vision encoder. To improve semantic alignment, learnable text prompts are employed to link visual features. Furthermore, SigLIP loss is applied to effectively handle the many-to-one relationship between images and unpaired text prompts, showcasing its adaptation in the medical field for the first time. Our approach is validated on multiple modalities, demonstrating superior performance over existing methods for AC and AS, in both same-dataset and cross-dataset evaluations. Unlike prior work, it does not rely on synthetic data or memory banks, and an ablation study confirms the contribution of each component. The code is available at https://github.com/mahshid1998/MadCLIP.

Dario Serez, Marco Cristani, Alessio Del Bue et al.

Attackers can deliberately perturb classifiers' input with subtle noise, altering final predictions. Among proposed countermeasures, adversarial purification employs generative networks to preprocess input images, filtering out adversarial noise. In this study, we propose specific generators, defined Multiple Latent Variable Generative Models (MLVGMs), for adversarial purification. These models possess multiple latent variables that naturally disentangle coarse from fine features. Taking advantage of these properties, we autoencode images to maintain class-relevant information, while discarding and re-sampling any detail, including adversarial noise. The procedure is completely training-free, exploring the generalization abilities of pre-trained MLVGMs on the adversarial purification downstream task. Despite the lack of large models, trained on billions of samples, we show that smaller MLVGMs are already competitive with traditional methods, and can be used as foundation models. Official code released at https://github.com/SerezD/gen_adversarial.

Mahshid Shiri, Cigdem Beyan, Vittorio Murino

Medical anomaly detection (AD) is challenging due to diverse imaging modalities, anatomical variations, and limited labeled data. We propose a novel approach combining visual adapters and prompt learning with Partial Optimal Transport (POT) and contrastive learning (CL) to improve CLIP's adaptability to medical images, particularly for AD. Unlike standard prompt learning, which often yields a single representation, our method employs multiple prompts aligned with local features via POT to capture subtle abnormalities. CL further enforces intra-class cohesion and inter-class separation. Our method achieves state-of-the-art results in few-shot, zero-shot, and cross-dataset scenarios without synthetic data or memory banks. The code is available at https://github.com/mahshid1998/MADPOT.

Massimiliano Ciranni, Vito Paolo Pastore, Roberto Di Via et al.

Deep learning model effectiveness in classification tasks is often challenged by the quality and quantity of training data whenever they are affected by strong spurious correlations between specific attributes and target labels. This results in a form of bias affecting training data, which typically leads to unrecoverable weak generalization in prediction. This paper aims at facing this problem by leveraging bias amplification with generated synthetic data: we introduce Diffusing DeBias (DDB), a novel approach acting as a plug-in for common methods of unsupervised model debiasing exploiting the inherent bias-learning tendency of diffusion models in data generation. Specifically, our approach adopts conditional diffusion models to generate synthetic bias-aligned images, which replace the original training set for learning an effective bias amplifier model that we subsequently incorporate into an end-to-end and a two-step unsupervised debiasing approach. By tackling the fundamental issue of bias-conflicting training samples memorization in learning auxiliary models, typical of this type of techniques, our proposed method beats current state-of-the-art in multiple benchmark datasets, demonstrating its potential as a versatile and effective tool for tackling bias in deep learning models. Code is available at https://github.com/Malga-Vision/DiffusingDeBias

Dario Serez, Marco Cristani, Alessio Del Bue et al.

In image generation, Multiple Latent Variable Generative Models (MLVGMs) employ multiple latent variables to gradually shape the final images, from global characteristics to finer and local details (e.g., StyleGAN, NVAE), emerging as powerful tools for diverse applications. Yet their generative dynamics remain only empirically observed, without a systematic understanding of each latent variable's impact. In this work, we propose a novel framework that quantifies the contribution of each latent variable using Mutual Information (MI) as a metric. Our analysis reveals that current MLVGMs often underutilize some latent variables, and provides actionable insights for their use in downstream applications. With this foundation, we introduce a method for generating synthetic data for Self-Supervised Contrastive Representation Learning (SSCRL). By leveraging the hierarchical and disentangled variables of MLVGMs, our approach produces diverse and semantically meaningful views without the need for real image data. Additionally, we introduce a Continuous Sampling (CS) strategy, where the generator dynamically creates new samples during SSCRL training, greatly increasing data variability. Our comprehensive experiments demonstrate the effectiveness of these contributions, showing that MLVGMs' generated views compete on par with or even surpass views generated from real data. This work establishes a principled approach to understanding and exploiting MLVGMs, advancing both generative modeling and self-supervised learning. Code and pre-trained models at: https://github.com/SerezD/mi_ml_gen.

Sebastiano Vascon, Samuel Rota Bulò, Vittorio Murino et al.

DSLib is an open-source implementation of the Dominant Set (DS) clustering algorithm written entirely in Matlab. The DS method is a graph-based clustering technique rooted in the evolutionary game theory that starts gaining lots of interest in the computer science community. Thanks to its duality with game theory and its strict relation to the notion of maximal clique, has been explored in several directions not only related to clustering problems. Applications in graph matching, segmentation, classification and medical imaging are common in literature. This package provides an implementation of the original DS clustering algorithm since no code has been officially released yet, together with a still growing collection of methods and variants related to it. Our library is integrable into a Matlab pipeline without dependencies, it is simple to use and easily extendable for upcoming works. The latest source code, the documentation and some examples can be downloaded from https://xwasco.github.io/DominantSetLibrary.

Chunlei Liu, Wenrui Ding, Jinyu Yang et al.

Small object tracking becomes an increasingly important task, which however has been largely unexplored in computer vision. The great challenges stem from the facts that: 1) small objects show extreme vague and variable appearances, and 2) they tend to be lost easier as compared to normal-sized ones due to the shaking of lens. In this paper, we propose a novel aggregation signature suitable for small object tracking, especially aiming for the challenge of sudden and large drift. We make three-fold contributions in this work. First, technically, we propose a new descriptor, named aggregation signature, based on saliency, able to represent highly distinctive features for small objects. Second, theoretically, we prove that the proposed signature matches the foreground object more accurately with a high probability. Third, experimentally, the aggregation signature achieves a high performance on multiple datasets, outperforming the state-of-the-art methods by large margins. Moreover, we contribute with two newly collected benchmark datasets, i.e., small90 and small112, for visually small object tracking. The datasets will be available in https://github.com/bczhangbczhang/.

Nuno Garcia, Pietro Morerio, Vittorio Murino

Diverse input data modalities can provide complementary cues for several tasks, usually leading to more robust algorithms and better performance. However, while a (training) dataset could be accurately designed to include a variety of sensory inputs, it is often the case that not all modalities could be available in real life (testing) scenarios, where a model has to be deployed. This raises the challenge of how to learn robust representations leveraging multimodal data in the training stage, while considering limitations at test time, such as noisy or missing modalities. This paper presents a new approach for multimodal video action recognition, developed within the unified frameworks of distillation and privileged information, named generalized distillation. Particularly, we consider the case of learning representations from depth and RGB videos, while relying on RGB data only at test time. We propose a new approach to train an hallucination network that learns to distill depth features through multiplicative connections of spatiotemporal representations, leveraging soft labels and hard labels, as well as distance between feature maps. We report state-of-the-art results on video action classification on the largest multimodal dataset available for this task, the NTU RGB+D. Code available at https://github.com/ncgarcia/modality-distillation .

Caterina Gallegati, Monica Bianchini, Franco Scarselli et al.

As generative models achieve unprecedented visual quality, the gold standard for image evaluation remains traditional feature-distribution metrics (e.g., FID). However, these metrics are provably hindered by the closed-vocabulary bottleneck of outdated features and the assumptive bias of rigid parametric formulations. Recent alternatives exploit modern backbones to solve the feature bottleneck, yet continue to suffer from parametric limitations. To close this gap, we introduce APEX (Assumption-free Projection-based Embedding eXamination), a novel evaluation framework leveraging the Sliced Wasserstein Distance as a mathematically grounded, assumption-free similarity measure. APEX inherits effective scalability to high-dimensional spaces, as we prove with theoretical and empirical evidences. Moreover, APEX is embedding-agnostic and uses two open-vocabulary foundation models, CLIP and DINOv2, as feature extractors. Benchmarking APEX against established baselines reveals superior robustness to visual degradations. Additionally, we show that APEX metrics exhibit intra- and cross-dataset stability, ensuring highly stable evaluations on out-of-domain datasets.

Yi Wei, Xuan Qi, Furao shen et al.

Piecewise affine neural networks (PANNs) provide a principled geometric perspective on neural network expressivity by characterizing the input--output map as a continuous piecewise affine (CPA) function whose complexity is governed by the number, arrangement, and shapes of its affine regions. However, existing interpretability and expressivity analyses often rely on indirect proxies (e.g., activation statistics or theoretical upper bounds) and rarely offer practical, accurate tools for enumerating and visualizing the induced region partition under realistic architectures and bounded input domains. In this work, we present AffineLens, a unified framework for computing the hyperplane arrangements and polyhedral structures underlying PANNs. Given a calibrated (bounded) input polytope, AffineLens identifies the subset of neuron-induced hyperplanes that intersect the domain, enumerates the resulting affine sub-regions in a layer-wise manner, and returns provably non-empty maximal CPA regions together with interior representatives. The framework further provides visualizations of region partitioning and decision boundaries, enabling qualitative inspection alongside quantitative region counts. By exploiting the affine restriction property of CPA networks under fixed activation patterns, AffineLens supports a broad class of modern components, including batch normalization, pooling, residual connections, multilayer perceptrons, and convolutional layers. Finally, we use AffineLens to perform a systematic empirical study of architectural expressivity, comparing networks through region complexity metrics and revealing how design choices influence the geometry of learned functions.

Xuan Qi, Yi Wei, Fanqi Yu et al.

Batch normalization (BN) is central to modern deep networks, but its effect on the realized function during training remains less understood than its optimization benefits. We study training-time BN in continuous piecewise-affine (CPA) networks through the geometry of switching hyperplanes and the induced affine-region partition. Conditioned on a mini-batch, we show that BN defines for each neuron a reference hyperplane through the batch centroid, and that breakpoint-switching hyperplanes are parallel translates whose offsets are expressed in batch-standardized coordinates and are independent of the raw bias. This yields an exact criterion for when a switching hyperplane intersects a local $\ell_\infty$ window and motivates a local region-density functional based on exact affine-region counts. Under explicit sufficient conditions, we show that BN increases expected local partition refinement in ReLU and more general piecewise-affine networks, and that this mechanism transfers locally through depth inside parent affine regions where the upstream representation map is an affine embedding. These results provide a function-level geometric account of training-time BN as a batch-conditional recentering mechanism near the data.

Sanket Thakur, Cigdem Beyan, Pietro Morerio et al.

In this technical report, we describe the Guided-Attention mechanism based solution for the short-term anticipation (STA) challenge for the EGO4D challenge. It combines the object detections, and the spatiotemporal features extracted from video clips, enhancing the motion and contextual information, and further decoding the object-centric and motion-centric information to address the problem of STA in egocentric videos. For the challenge, we build our model on top of StillFast with Guided Attention applied on fast network. Our model obtains better performance on the validation set and also achieves state-of-the-art (SOTA) results on the challenge test set for EGO4D Short-Term Object Interaction Anticipation Challenge.

Sanket Thakur, Cigdem Beyan, Pietro Morerio et al.

Short-term action anticipation (STA) in first-person videos is a challenging task that involves understanding the next active object interactions and predicting future actions. Existing action anticipation methods have primarily focused on utilizing features extracted from video clips, but often overlooked the importance of objects and their interactions. To this end, we propose a novel approach that applies a guided attention mechanism between the objects, and the spatiotemporal features extracted from video clips, enhancing the motion and contextual information, and further decoding the object-centric and motion-centric information to address the problem of STA in egocentric videos. Our method, GANO (Guided Attention for Next active Objects) is a multi-modal, end-to-end, single transformer-based network. The experimental results performed on the largest egocentric dataset demonstrate that GANO outperforms the existing state-of-the-art methods for the prediction of the next active object label, its bounding box location, the corresponding future action, and the time to contact the object. The ablation study shows the positive contribution of the guided attention mechanism compared to other fusion methods. Moreover, it is possible to improve the next active object location and class label prediction results of GANO by just appending the learnable object tokens with the region of interest embeddings.

Julio Ivan Davila Carrazco, Suvarna Kishorkumar Kadam, Pietro Morerio et al.

In this paper, we introduce a novel framework for the challenging problem of One-Shot Unsupervised Domain Adaptation (OSUDA), which aims to adapt to a target domain with only a single unlabeled target sample. Unlike existing approaches that rely on large labeled source and unlabeled target data, our Target-driven One-Shot UDA (TOS-UDA) approach employs a learnable augmentation strategy guided by the target sample's style to align the source distribution with the target distribution. Our method consists of three modules: an augmentation module, a style alignment module, and a classifier. Unlike existing methods, our augmentation module allows for strong transformations of the source samples, and the style of the single target sample available is exploited to guide the augmentation by ensuring perceptual similarity. Furthermore, our approach integrates augmentation with style alignment, eliminating the need for separate pre-training on additional datasets. Our method outperforms or performs comparably to existing OS-UDA methods on the Digits and DomainNet benchmarks.

Waqar Ahmed, Andrea Zunino, Pietro Morerio et al.

The concept of compressing deep Convolutional Neural Networks (CNNs) is essential to use limited computation, power, and memory resources on embedded devices. However, existing methods achieve this objective at the cost of a drop in inference accuracy in computer vision tasks. To address such a drawback, we propose a framework that leverages knowledge distillation along with customizable block-wise optimization to learn a lightweight CNN structure while preserving better control over the compression-performance tradeoff. Considering specific resource constraints, e.g., floating-point operations per inference (FLOPs) or model-parameters, our method results in a state of the art network compression while being capable of achieving better inference accuracy. In a comprehensive evaluation, we demonstrate that our method is effective, robust, and consistent with results over a variety of network architectures and datasets, at negligible training overhead. In particular, for the already compact network MobileNet_v2, our method offers up to 2x and 5.2x better model compression in terms of FLOPs and model-parameters, respectively, while getting 1.05% better model performance than the baseline network.

Waqar Ahmed, Pietro Morerio, Vittorio Murino

The majority of existing Unsupervised Domain Adaptation (UDA) methods presumes source and target domain data to be simultaneously available during training. Such an assumption may not hold in practice, as source data is often inaccessible (e.g., due to privacy reasons). On the contrary, a pre-trained source model is always considered to be available, even though performing poorly on target due to the well-known domain shift problem. This translates into a significant amount of misclassifications, which can be interpreted as structured noise affecting the inferred target pseudo-labels. In this work, we cast UDA as a pseudo-label refinery problem in the challenging source-free scenario. We propose a unified method to tackle adaptive noise filtering and pseudo-label refinement. A novel Negative Ensemble Learning technique is devised to specifically address noise in pseudo-labels, by enhancing diversity in ensemble members with different stochastic (i) input augmentation and (ii) feedback. In particular, the latter is achieved by leveraging the novel concept of Disjoint Residual Labels, which allow diverse information to be fed to the different members. A single target model is eventually trained with the refined pseudo-labels, which leads to a robust performance on the target domain. Extensive experiments show that the proposed method, named Adaptive Pseudo-Label Refinement, achieves state-of-the-art performance on major UDA benchmarks, such as Digit5, PACS, Visda-C, and DomainNet, without using source data at all.

Federico Marmoreo, Julio Ivan Davila Carrazco, Vittorio Murino et al.

In Generalized Zero-Shot Learning (GZSL), unseen categories (for which no visual data are available at training time) can be predicted by leveraging their class embeddings (e.g., a list of attributes describing them) together with a complementary pool of seen classes (paired with both visual data and class embeddings). Despite GZSL is arguably challenging, we posit that knowing in advance the class embeddings, especially for unseen categories, is an actual limit of the applicability of GZSL towards real-world scenarios. To relax this assumption, we propose Open Zero-Shot Learning (OZSL) to extend GZSL towards the open-world settings. We formalize OZSL as the problem of recognizing seen and unseen classes (as in GZSL) while also rejecting instances from unknown categories, for which neither visual data nor class embeddings are provided. We formalize the OZSL problem introducing evaluation protocols, error metrics and benchmark datasets. We also suggest to tackle the OZSL problem by proposing the idea of performing unknown feature generation (instead of only unseen features generation as done in GZSL). We achieve this by optimizing a generative process to sample unknown class embeddings as complementary to the seen and the unseen. We intend these results to be the ground to foster future research, extending the standard closed-world zero-shot learning (GZSL) with the novel open-world counterpart (OZSL).

Federico Marmoreo, Jacopo Cavazza, Vittorio Murino

In this paper, we address zero-shot learning (ZSL), the problem of recognizing categories for which no labeled visual data are available during training. We focus on the transductive setting, in which unlabelled visual data from unseen classes is available. State-of-the-art paradigms in ZSL typically exploit generative adversarial networks to synthesize visual features from semantic attributes. We posit that the main limitation of these approaches is to adopt a single model to face two problems: 1) generating realistic visual features, and 2) translating semantic attributes into visual cues. Differently, we propose to decouple such tasks, solving them separately. In particular, we train an unconditional generator to solely capture the complexity of the distribution of visual data and we subsequently pair it with a conditional generator devoted to enrich the prior knowledge of the data distribution with the semantic content of the class embeddings. We present a detailed ablation study to dissect the effect of our proposed decoupling approach, while demonstrating its superiority over the related state-of-the-art.

Milind G. Padalkar, Carlos Beltrán-González, Matteo Bustreo et al.

This paper presents a novel setup for automatic visual inspection of cracks in ceramic tile as well as studies the effect of various classifiers and height-varying illumination conditions for this task. The intuition behind this setup is that cracks can be better visualized under specific lighting conditions than others. Our setup, which is designed for field work with constraints in its maximum dimensions, can acquire images for crack detection with multiple lighting conditions using the illumination sources placed at multiple heights. Crack detection is then performed by classifying patches extracted from the acquired images in a sliding window fashion. We study the effect of lights placed at various heights by training classifiers both on customized as well as state-of-the-art architectures and evaluate their performance both at patch-level and image-level, demonstrating the effectiveness of our setup. More importantly, ours is the first study that demonstrates how height-varying illumination conditions can affect crack detection with the use of existing state-of-the-art classifiers. We provide an insight about the illumination conditions that can help in improving crack detection in a challenging real-world industrial environment.

Marco Cristani, Alessio Del Bue, Vittorio Murino et al.

One of the main and most effective measures to contain the recent viral outbreak is the maintenance of the so-called Social Distancing (SD). To comply with this constraint, workplaces, public institutions, transports and schools will likely adopt restrictions over the minimum inter-personal distance between people. Given this actual scenario, it is crucial to massively measure the compliance to such physical constraint in our life, in order to figure out the reasons of the possible breaks of such distance limitations, and understand if this implies a possible threat given the scene context. All of this, complying with privacy policies and making the measurement acceptable. To this end, we introduce the Visual Social Distancing (VSD) problem, defined as the automatic estimation of the inter-personal distance from an image, and the characterization of the related people aggregations. VSD is pivotal for a non-invasive analysis to whether people comply with the SD restriction, and to provide statistics about the level of safety of specific areas whenever this constraint is violated. We then discuss how VSD relates with previous literature in Social Signal Processing and indicate which existing Computer Vision methods can be used to manage such problem. We conclude with future challenges related to the effectiveness of VSD systems, ethical implications and future application scenarios.

Maya Aghaei, Matteo Bustreo, Pietro Morerio et al.

The design of an automatic visual inspection system is usually performed in two stages. While the first stage consists in selecting the most suitable hardware setup for highlighting most effectively the defects on the surface to be inspected, the second stage concerns the development of algorithmic solutions to exploit the potentials offered by the collected data. In this paper, first, we present a novel illumination setup embedding four illumination configurations to resemble diffused, dark-field, and front lighting techniques. Second, we analyze the contributions brought by deploying the proposed setup in training phase only - mimicking the scenario in which an already developed visual inspection system cannot be modified on the customer site - and in evaluation phase. Along with an exhaustive set of experiments, in this paper, we demonstrate the suitability of the proposed setup for effective illumination of complex-objects, defined as manufactured items with variable surface characteristics that cannot be determined a priori. Moreover, we discuss the importance of multiple light configurations availability during training and their natural boosting effect which, without the need to modify the system design in evaluation phase, lead to improvements in the overall system performance.

Avik Hati, Matteo Bustreo, Diego Sona et al.

In this paper, we tackle the task of automatically analyzing 3D volumetric scans obtained from computed tomography (CT) devices. In particular, we address a particular task for which data is very limited: the segmentation of ancient Egyptian mummies CT scans. We aim at digitally unwrapping the mummy and identify different segments such as body, bandages and jewelry. The problem is complex because of the lack of annotated data for the different semantic regions to segment, thus discouraging the use of strongly supervised approaches. We, therefore, propose a weakly supervised and efficient interactive segmentation method to solve this challenging problem. After segmenting the wrapped mummy from its exterior region using histogram analysis and template matching, we first design a voxel distance measure to find an approximate solution for the body and bandage segments. Here, we use geodesic distances since voxel features as well as spatial relationship among voxels is incorporated in this measure. Next, we refine the solution using a GrabCut based segmentation together with a tracking method on the slices of the scan that assigns labels to different regions in the volume, using limited supervision in the form of scribbles drawn by the user. The efficiency of the proposed method is demonstrated using visualizations and validated through quantitative measures and qualitative unwrapping of the mummy.

Andrea Zunino, Sarah Adel Bargal, Riccardo Volpi et al.

Conventionally, AI models are thought to trade off explainability for lower accuracy. We develop a training strategy that not only leads to a more explainable AI system for object classification, but as a consequence, suffers no perceptible accuracy degradation. Explanations are defined as regions of visual evidence upon which a deep classification network makes a decision. This is represented in the form of a saliency map conveying how much each pixel contributed to the network's decision. Our training strategy enforces a periodic saliency-based feedback to encourage the model to focus on the image regions that directly correspond to the ground-truth object. We quantify explainability using an automated metric, and using human judgement. We propose explainability as a means for bridging the visual-semantic gap between different domains where model explanations are used as a means of disentagling domain specific information from otherwise relevant features. We demonstrate that this leads to improved generalization to new domains without hindering performance on the original domain.

Ruggero Ragonesi, Riccardo Volpi, Jacopo Cavazza et al.

We are interested in learning data-driven representations that can generalize well, even when trained on inherently biased data. In particular, we face the case where some attributes (bias) of the data, if learned by the model, can severely compromise its generalization properties. We tackle this problem through the lens of information theory, leveraging recent findings for a differentiable estimation of mutual information. We propose a novel end-to-end optimization strategy, which simultaneously estimates and minimizes the mutual information between the learned representation and the data attributes. When applied on standard benchmarks, our model shows comparable or superior classification performance with respect to state-of-the-art approaches. Moreover, our method is general enough to be applicable to the problem of ``algorithmic fairness'', with competitive results.

Xiangping Zhu, Xiatian Zhu, Minxian Li et al.

Existing person re-identification (re-id) methods mostly exploit a large set of cross-camera identity labelled training data. This requires a tedious data collection and annotation process, leading to poor scalability in practical re-id applications. On the other hand unsupervised re-id methods do not need identity label information, but they usually suffer from much inferior and insufficient model performance. To overcome these fundamental limitations, we propose a novel person re-identification paradigm based on an idea of independent per-camera identity annotation. This eliminates the most time-consuming and tedious inter-camera identity labelling process, significantly reducing the amount of human annotation efforts. Consequently, it gives rise to a more scalable and more feasible setting, which we call Intra-Camera Supervised (ICS) person re-id, for which we formulate a Multi-tAsk mulTi-labEl (MATE) deep learning method. Specifically, MATE is designed for self-discovering the cross-camera identity correspondence in a per-camera multi-task inference framework. Extensive experiments demonstrate the cost-effectiveness superiority of our method over the alternative approaches on three large person re-id datasets. For example, MATE yields 88.7% rank-1 score on Market-1501 in the proposed ICS person re-id setting, significantly outperforming unsupervised learning models and closely approaching conventional fully supervised learning competitors.

Pietro Morerio, Riccardo Volpi, Ruggero Ragonesi et al.

We investigate and characterize the inherent resilience of conditional Generative Adversarial Networks (cGANs) against noise in their conditioning labels, and exploit this fact in the context of Unsupervised Domain Adaptation (UDA). In UDA, a classifier trained on the labelled source set can be used to infer pseudo-labels on the unlabelled target set. However, this will result in a significant amount of misclassified examples (due to the well-known domain shift issue), which can be interpreted as noise injection in the ground-truth labels for the target set. We show that cGANs are, to some extent, robust against such "shift noise". Indeed, cGANs trained with noisy pseudo-labels, are able to filter such noise and generate cleaner target samples. We exploit this finding in an iterative procedure where a generative model and a classifier are jointly trained: in turn, the generator allows to sample cleaner data from the target distribution, and the classifier allows to associate better labels to target samples, progressively refining target pseudo-labels. Results on common benchmarks show that our method performs better or comparably with the unsupervised domain adaptation state of the art.

Nuno C. Garcia, Sarah Adel Bargal, Vitaly Ablavsky et al.

In this work, we address the problem of learning an ensemble of specialist networks using multimodal data, while considering the realistic and challenging scenario of possible missing modalities at test time. Our goal is to leverage the complementary information of multiple modalities to the benefit of the ensemble and each individual network. We introduce a novel Distillation Multiple Choice Learning framework for multimodal data, where different modality networks learn in a cooperative setting from scratch, strengthening one another. The modality networks learned using our method achieve significantly higher accuracy than if trained separately, due to the guidance of other modalities. We evaluate this approach on three video action recognition benchmark datasets. We obtain state-of-the-art results in comparison to other approaches that work with missing modalities at test time.

Shahab Aslani, Vittorio Murino, Michael Dayan et al.

This paper presents a simple and effective generalization method for magnetic resonance imaging (MRI) segmentation when data is collected from multiple MRI scanning sites and as a consequence is affected by (site-)domain shifts. We propose to integrate a traditional encoder-decoder network with a regularization network. This added network includes an auxiliary loss term which is responsible for the reduction of the domain shift problem and for the resulting improved generalization. The proposed method was evaluated on multiple sclerosis lesion segmentation from MRI data. We tested the proposed model on an in-house clinical dataset including 117 patients from 56 different scanning sites. In the experiments, our method showed better generalization performance than other baseline networks.

Xiangping Zhu, Pietro Morerio, Vittorio Murino

Pedestrian attributes, e.g., hair length, clothes type and color, locally describe the semantic appearance of a person. Training person re-identification (ReID) algorithms under the supervision of such attributes have proven to be effective in extracting local features which are important for ReID. Unlike person identity, attributes are consistent across different domains (or datasets). However, most of ReID datasets lack attribute annotations. On the other hand, there are several datasets labeled with sufficient attributes for the case of pedestrian attribute recognition. Exploiting such data for ReID purpose can be a way to alleviate the shortage of attribute annotations in ReID case. In this work, an unsupervised domain adaptive ReID feature learning framework is proposed to make full use of attribute annotations. We propose to transfer attribute-related features from their original domain to the ReID one: to this end, we introduce an adversarial discriminative domain adaptation method in order to learn domain invariant features for encoding semantic attributes. Experiments on three large-scale datasets validate the effectiveness of the proposed ReID framework.

Xiangping Zhu, Xiatian Zhu, Minxian Li et al.

Existing person re-identification (re-id) methods rely mostly on a large set of inter-camera identity labelled training data, requiring a tedious data collection and annotation process therefore leading to poor scalability in practical re-id applications. To overcome this fundamental limitation, we consider person re-identification without inter-camera identity association but only with identity labels independently annotated within each individual camera-view. This eliminates the most time-consuming and tedious inter-camera identity labelling process in order to significantly reduce the amount of human efforts required during annotation. It hence gives rise to a more scalable and more feasible learning scenario, which we call Intra-Camera Supervised (ICS) person re-id. Under this ICS setting with weaker label supervision, we formulate a Multi-Task Multi-Label (MTML) deep learning method. Given no inter-camera association, MTML is specially designed for self-discovering the inter-camera identity correspondence. This is achieved by inter-camera multi-label learning under a joint multi-task inference framework. In addition, MTML can also efficiently learn the discriminative re-id feature representations by fully using the available identity labels within each camera-view. Extensive experiments demonstrate the performance superiority of our MTML model over the state-of-the-art alternative methods on three large-scale person re-id datasets in the proposed intra-camera supervised learning setting.

Andrés F. Pérez, Valentina Sanguineti, Pietro Morerio et al.

In this paper, we investigate how to learn rich and robust feature representations for audio classification from visual data and acoustic images, a novel audio data modality. Former models learn audio representations from raw signals or spectral data acquired by a single microphone, with remarkable results in classification and retrieval. However, such representations are not so robust towards variable environmental sound conditions. We tackle this drawback by exploiting a new multimodal labeled action recognition dataset acquired by a hybrid audio-visual sensor that provides RGB video, raw audio signals, and spatialized acoustic data, also known as acoustic images, where the visual and acoustic images are aligned in space and synchronized in time. Using this richer information, we train audio deep learning models in a teacher-student fashion. In particular, we distill knowledge into audio networks from both visual and acoustic image teachers. Our experiments suggest that the learned representations are more powerful and have better generalization capabilities than the features learned from models trained using just single-microphone audio data.

Riccardo Volpi, Vittorio Murino

We are concerned with the vulnerability of computer vision models to distributional shifts. We formulate a combinatorial optimization problem that allows evaluating the regions in the image space where a given model is more vulnerable, in terms of image transformations applied to the input, and face it with standard search algorithms. We further embed this idea in a training procedure, where we define new data augmentation rules according to the image transformations that the current model is most vulnerable to, over iterations. An empirical evaluation on classification and semantic segmentation problems suggests that the devised algorithm allows to train models that are more robust against content-preserving image manipulations and, in general, against distributional shifts.

Sarah Adel Bargal, Andrea Zunino, Vitali Petsiuk et al.

We propose Guided Zoom, an approach that utilizes spatial grounding of a model's decision to make more informed predictions. It does so by making sure the model has "the right reasons" for a prediction, defined as reasons that are coherent with those used to make similar correct decisions at training time. The reason/evidence upon which a deep convolutional neural network makes a prediction is defined to be the spatial grounding, in the pixel space, for a specific class conditional probability in the model output. Guided Zoom examines how reasonable such evidence is for each of the top-k predicted classes, rather than solely trusting the top-1 prediction. We show that Guided Zoom improves the classification accuracy of a deep convolutional neural network model and obtains state-of-the-art results on three fine-grained classification benchmark datasets.

Shahab Aslani, Michael Dayan, Loredana Storelli et al.

In this paper, we present an automated approach for segmenting multiple sclerosis (MS) lesions from multi-modal brain magnetic resonance images. Our method is based on a deep end-to-end 2D convolutional neural network (CNN) for slice-based segmentation of 3D volumetric data. The proposed CNN includes a multi-branch downsampling path, which enables the network to encode information from multiple modalities separately. Multi-scale feature fusion blocks are proposed to combine feature maps from different modalities at different stages of the network. Then, multi-scale feature upsampling blocks are introduced to upsize combined feature maps to leverage information from lesion shape and location. We trained and tested the proposed model using orthogonal plane orientations of each 3D modality to exploit the contextual information in all directions. The proposed pipeline is evaluated on two different datasets: a private dataset including 37 MS patients and a publicly available dataset known as the ISBI 2015 longitudinal MS lesion segmentation challenge dataset, consisting of 14 MS patients. Considering the ISBI challenge, at the time of submission, our method was amongst the top performing solutions. On the private dataset, using the same array of performance metrics as in the ISBI challenge, the proposed approach shows high improvements in MS lesion segmentation compared with other publicly available tools.

Nuno C. Garcia, Pietro Morerio, Vittorio Murino

Heterogeneous data modalities can provide complementary cues for several tasks, usually leading to more robust algorithms and better performance. However, while training data can be accurately collected to include a variety of sensory modalities, it is often the case that not all of them are available in real life (testing) scenarios, where a model has to be deployed. This raises the challenge of how to extract information from multimodal data in the training stage, in a form that can be exploited at test time, considering limitations such as noisy or missing modalities. This paper presents a new approach in this direction for RGB-D vision tasks, developed within the adversarial learning and privileged information frameworks. We consider the practical case of learning representations from depth and RGB videos, while relying only on RGB data at test time. We propose a new approach to train a hallucination network that learns to distill depth information via adversarial learning, resulting in a clean approach without several losses to balance or hyperparameters. We report state-of-the-art results on object classification on the NYUD dataset and video action recognition on the largest multimodal dataset available for this task, the NTU RGB+D, as well as on the Northwestern-UCLA.

Riccardo Volpi, Hongseok Namkoong, Ozan Sener et al.

We are concerned with learning models that generalize well to different \emph{unseen} domains. We consider a worst-case formulation over data distributions that are near the source domain in the feature space. Only using training data from a single source distribution, we propose an iterative procedure that augments the dataset with examples from a fictitious target domain that is "hard" under the current model. We show that our iterative scheme is an adaptive data augmentation method where we append adversarial examples at each iteration. For softmax losses, we show that our method is a data-dependent regularization scheme that behaves differently from classical regularizers that regularize towards zero (e.g., ridge or lasso). On digit recognition and semantic segmentation tasks, our method learns models improve performance across a range of a priori unknown target domains.

Andrea Zunino, Sarah Adel Bargal, Pietro Morerio et al.

We propose a guided dropout regularizer for deep networks based on the evidence of a network prediction defined as the firing of neurons in specific paths. In this work, we utilize the evidence at each neuron to determine the probability of dropout, rather than dropping out neurons uniformly at random as in standard dropout. In essence, we dropout with higher probability those neurons which contribute more to decision making at training time. This approach penalizes high saliency neurons that are most relevant for model prediction, i.e. those having stronger evidence. By dropping such high-saliency neurons, the network is forced to learn alternative paths in order to maintain loss minimization, resulting in a plasticity-like behavior, a characteristic of human brains too. We demonstrate better generalization ability, an increased utilization of network neurons, and a higher resilience to network compression using several metrics over four image/video recognition benchmarks.

Jacopo Cavazza, Pietro Morerio, Vittorio Murino

Despite the recent deep learning (DL) revolution, kernel machines still remain powerful methods for action recognition. DL has brought the use of large datasets and this is typically a problem for kernel approaches, which are not scaling up efficiently due to kernel Gram matrices. Nevertheless, kernel methods are still attractive and more generally applicable since they can equally manage different sizes of the datasets, also in cases where DL techniques show some limitations. This work investigates these issues by proposing an explicit approximated representation that, together with a linear model, is an equivalent, yet scalable, implementation of a kernel machine. Our approximation is directly inspired by the exact feature map that is induced by an RBF Gaussian kernel but, unlike the latter, it is finite dimensional and very compact. We justify the soundness of our idea with a theoretical analysis which proves the unbiasedness of the approximation, and provides a vanishing bound for its variance, which is shown to decrease much rapidly than in alternative methods in the literature. In a broad experimental validation, we assess the superiority of our approximation in terms of 1) ease and speed of training, 2) compactness of the model, and 3) improvements with respect to the state-of-the-art performance.