The fine-tuning of deep pre-trained models has revealed compositional properties, with multiple specialized modules that can be arbitrarily composed into a single, multi-task model. However, identifying the conditions that promote compositionality remains an open issue, with recent efforts concentrating mainly on linearized networks. We conduct a theoretical study that attempts to demystify compositionality in standard non-linear networks through the second-order Taylor approximation of the loss function. The proposed formulation highlights the importance of staying within the pre-training basin to achieve composable modules. Moreover, it provides the basis for two dual incremental training algorithms: the one from the perspective of multiple models trained individually, while the other aims to optimize the composed model as a whole. We probe their application in incremental classification tasks and highlight some valuable skills. In fact, the pool of incrementally learned modules not only supports the creation of an effective multi-task model but also enables unlearning and specialization in certain tasks. Code available at <https://github.com/aimagelab/mammoth>

Model merging has emerged as a crucial technique in Deep Learning, enabling the integration of multiple models into a unified system while preserving performance and scalability. In this respect, the compositional properties of low-rank adaptation techniques (e.g., LoRA) have proven beneficial, as simple averaging LoRA modules yields a single model that mostly integrates the capabilities of all individual modules. Building on LoRA, we take a step further by imposing that the merged model matches the responses of all learned modules. Solving this ob- jective in closed form yields an indeterminate system with A and B as unknown variables, indicating the existence of infinitely many closed-form solutions. To address this challenge, we introduce LoRM, an alternating optimization strategy that trains one LoRA matrix at a time. This allows solving for each unknown variable individually, thus finding a unique solution. We apply our proposed methodology to Federated Class-Incremental Learning (FCIL), ensuring alignment of model responses both between clients and across tasks. Our method demonstrates state-of-the-art performance across a range of FCIL scenarios. The code to reproduce our experiments is available at github.com/aimagelab/fed-mammoth.

Foundation models serve as the backbone for numerous specialized models developed through fine-tuning. However, when the underlying pretrained model is updated or retrained (e.g., on larger and more curated datasets), the fine-tuned model becomes obsolete, losing its utility and requiring retraining. This raises the question: is it possible to transfer fine-tuning to a new release of the model? In this work, we investigate how to transfer fine-tuning to a new checkpoint without having to re-train, in a data-free manner. To do so, we draw principles from model re-basin and provide a recipe based on weight permutations to re-base the modifications made to the original base model, often called task vector. In particular, our approach tailors model re-basin for Transformer models, taking into account the challenges of residual connections and multi-head attention layers. Specifically, we propose a two-level method rooted in spectral theory, initially permuting the attention heads and subsequently adjusting parameters within select pairs of heads. Through extensive experiments on visual and textual tasks, we achieve the seamless transfer of fine-tuned knowledge to new pre-trained backbones without relying on a single training step or datapoint. Code is available at https://github.com/aimagelab/TransFusion.

Open-Vocabulary object detectors can generalize to an unrestricted set of categories through simple textual prompting. However, adapting these models to rare classes or reinforcing their abilities on multiple specialized domains remains essential. While recent methods rely on monolithic adaptation strategies with a single set of weights, we embrace modular deep learning. We introduce DitHub, a framework designed to build and maintain a library of efficient adaptation modules. Inspired by Version Control Systems, DitHub manages expert modules as branches that can be fetched and merged as needed. This modular approach allows us to conduct an in-depth exploration of the compositional properties of adaptation modules, marking the first such study in Object Detection. Our method achieves state-of-the-art performance on the ODinW-13 benchmark and ODinW-O, a newly introduced benchmark designed to assess class reappearance.

In this paper, we address the challenges associated with merging low-rank adaptations of large neural networks. With the rise of parameter-efficient adaptation techniques, such as Low-Rank Adaptation (LoRA), model fine-tuning has become more accessible. While fine-tuning models with LoRA is highly efficient, existing merging methods often sacrifice this efficiency by merging fully-sized weight matrices. We propose the Core Space merging framework, which enables the merging of LoRA-adapted models within a common alignment basis, thereby preserving the efficiency of low-rank adaptation while substantially improving accuracy across tasks. We further provide a formal proof that projection into Core Space ensures no loss of information and provide a complexity analysis showing the efficiency gains. Extensive empirical results demonstrate that Core Space significantly improves existing merging techniques and achieves state-of-the-art results on both vision and language tasks while utilizing a fraction of the computational resources. Codebase is available at https://github.com/apanariello4/core-space-merging.

Continual Learning (CL) has emerged as a paramount area in Artificial Intelligence (AI) because of its ability to learn multiple tasks sequentially without significant performance degradation. Despite the growing interest in CL frameworks, a critical aspect must be addressed: the inherent biases within training data. In this work, we show that, if overlooked, these biases can significantly impair the efficacy of continual learning models by inducing reliance on suboptimal shortcuts during data stream and memory retention, exacerbating catastrophic forgetting. In response, we present Learning without Shortcuts (LwS), which sets forth two primary objectives: (i) to identify and mitigate the exploitation of spurious correlations within the data stream and (ii) to develop a novel mechanism that constructs a fair memory buffer used in replay-based CL strategies. Our buffer construction strategy exploits the model confidence in a given example to balance the portion of samples per class, hence their contribution when replay activates. Unlike existing methods, LwS is agnostic to protected attributes, and results highlight that the proposed solution is indeed resilient to spurious correlations in CL settings. Code is available at https://github.com/aimagelab/mammoth.

Per-object distance estimation is critical in surveillance and autonomous driving, where safety is crucial. While existing methods rely on geometric or deep supervised features, only a few attempts have been made to leverage self-supervised learning. In this respect, our paper draws inspiration from Masked Image Modeling (MiM) and extends it to multi-object tasks . While MiM focuses on extracting global image-level representations, it struggles with individual objects within the image. This is detrimental for distance estimation, as objects far away correspond to negligible portions of the image. Conversely, our strategy, termed Masked Object Modeling ( MoM ), enables a novel application of masking techniques. In a few words, we devise an auxiliary objective that reconstructs the portions of the image pertaining to the objects detected in the scene. The training phase is performed in a single unified stage, simultaneously optimizing the masking objective and the downstream loss ( i.e ., distance estimation). We evaluate the effectiveness of MoM on a novel reference architecture (DistFormer) on the standard KITTI, NuScenes, and MOTSynth datasets. Our evaluation reveals that our framework surpasses the SoTA and highlights its robust regularization properties. The MoM strategy enhances both zero-shot and few-shot capabilities, from synthetic to real domain. Finally, it furthers the robustness of the model in the presence of occluded or poorly detected objects. • We introduce a novel self-supervised approach called Masked Object Modeling (MoM). • The paper proposes a novel framework for distance estimation from monocular images. • We show improved transfer capabilities for zero-shot and fine-tuning when using MoM. • We reach state-of-the-art performance on several distance estimation benchmarks.

Facial Expression Recognition (FER) models based on deep learning mostly rely on a supervised train-once-test-all approach. These approaches assume that a model trained on an in-the-wild facial expression dataset with one type of domain distribution will perform well on a test dataset with a domain distribution shift. However, facial images in real-world can be from different domain distributions from which the model has been trained. However, re-training models on only new domain distributions will severely affect the performance of the previous domain. Re-training on all previous and new data can improve overall performance but is computationally expansive. In this study, we oppose the train-once-test-all approach and propose a buffer-based continual learning approach to enhance the performance of multiple in-the-wild datasets. We propose a model that continually leverages attention to important facial features from the pre-trained model to improve performance in multiple datasets. We validated our model using split-in-the-wild datasets where the dataset is provided to the model in an incremental setting instead of all at once. Furthermore, to evaluate the model performance, we continually used three in-the-wild datasets representing different domains (Domain-FER). Extensive experiments on these datasets reveal that the proposed model achieves better results than other Continual FER models.

We present Saliency-driven Experience Replay - SER - a biologically-plausible approach based on replicating human visual saliency to enhance classification models in continual learning settings. Inspired by neurophysiological evidence that the primary visual cortex does not contribute to object manifold untangling for categorization and that primordial saliency biases are still embedded in the modern brain, we propose to employ auxiliary saliency prediction features as a modulation signal to drive and stabilize the learning of a sequence of non-i.i.d. classification tasks. Experimental results confirm that SER effectively enhances the performance (in some cases up to about twenty percent points) of state-of-the-art continual learning methods, both in class-incremental and task-incremental settings. Moreover, we show that saliency-based modulation successfully encourages the learning of features that are more robust to the presence of spurious features and to adversarial attacks than baseline methods. Code is available at: https://github.com/perceivelab/SER

This work proposes a self-supervised training strategy designed for combinatorial problems. An obstacle in applying supervised paradigms to such problems is the need for costly target solutions often produced with exact solvers. Inspired by semi- and self-supervised learning, we show that generative models can be trained by sampling multiple solutions and using the best one according to the problem objective as a pseudo-label. In this way, we iteratively improve the model generation capability by relying only on its self-supervision, eliminating the need for optimality information. We validate this Self-Labeling Improvement Method (SLIM) on the Job Shop Scheduling (JSP), a complex combinatorial problem that is receiving much attention from the neural combinatorial community. We propose a generative model based on the well-known Pointer Network and train it with SLIM. Experiments on popular benchmarks demonstrate the potential of this approach as the resulting models outperform constructive heuristics and state-of-the-art learning proposals for the JSP. Lastly, we prove the robustness of SLIM to various parameters and its generality by applying it to the Traveling Salesman Problem.

End-to-end transformer-based trackers have achieved remarkable performance on most human-related datasets. However, training these trackers in heterogeneous scenarios poses significant challenges, including negative interference - where the model learns conflicting scene-specific parameters - and limited domain generalization, which often necessitates expensive fine-tuning to adapt the models to new domains. In response to these challenges, we introduce Parameter-efficient Scenario-specific Tracking Architecture (PASTA), a novel framework that combines Parameter-Efficient Fine-Tuning (PEFT) and Modular Deep Learning (MDL). Specifically, we define key scenario attributes (e.g, camera-viewpoint, lighting condition) and train specialized PEFT modules for each attribute. These expert modules are combined in parameter space, enabling systematic generalization to new domains without increasing inference time. Extensive experiments on MOTSynth, along with zero-shot evaluations on MOT17 and PersonPath22 demonstrate that a neural tracker built from carefully selected modules surpasses its monolithic counterpart. We release models and code.

The failures of Deep Networks can sometimes be ascribed to biases in the data or algorithmic choices. Existing debiasing approaches exploit prior knowledge to avoid unintended solutions; we acknowledge that, in real-world settings, it could be unfeasible to gather enough prior information to characterize the bias, or it could even raise ethical considerations. We hence propose a novel debiasing approach, termed ClusterFix, which does not require any external hint about the nature of biases. Such an approach alters the standard empirical risk minimization and introduces a per-example weight, encoding how critical and far from the majority an example is. Notably, the weights consider how difficult it is for the model to infer the correct pseudo-label, which is obtained in a self-supervised manner by dividing examples into multiple clusters. Extensive experiments show that the misclassification error incurred in identifying the correct cluster allows for identifying examples prone to bias-related issues. As a result, our approach outperforms existing methods on standard benchmarks for bias removal and fairness.

While biological intelligence grows organically as new knowledge is gathered throughout life, Artificial Neural Networks forget catastrophically whenever they face a changing training data distribution. Rehearsal-based Continual Learning (CL) approaches have been established as a versatile and reliable solution to overcome this limitation; however, sudden input disruptions and memory constraints are known to alter the consistency of their predictions. We study this phenomenon by investigating the geometric characteristics of the learner’s latent space and find that replayed data points of different classes increasingly mix up, interfering with classification. Hence, we propose a geometric regularizer that enforces weak requirements on the Laplacian spectrum of the latent space, promoting a partitioning behavior. Our proposal, called Continual Spectral Regularizer for Incremental Learning (CaSpeR-IL), can be easily combined with any rehearsal-based CL approach and improves the performance of SOTA methods on standard benchmarks.

The usage of Multi Instance Learning (MIL) for classifying Whole Slide Images (WSIs) has recently increased. Due to their gigapixel size, the pixel-level annotation of such data is extremely expensive and time-consuming, practically unfeasible. For this reason, multiple automatic approaches have been raised in the last years to support clinical practice and diagnosis. Unfortunately, most state-of-the-art proposals apply attention mechanisms without considering the spatial instance correlation and usually work on a single-scale resolution. To leverage the full potential of pyramidal structured WSI, we propose a graph-based multi-scale MIL approach, DAS-MIL. Our model comprises three modules: i) a self-supervised feature extractor, ii) a graph-based architecture that precedes the MIL mechanism and aims at creating a more contextualized representation of the WSI structure by considering the mutual (spatial) instance correlation both inter and intra-scale. Finally, iii) a (self) distillation loss between resolutions is introduced to compensate for their informative gap and significantly improve the final prediction. The effectiveness of the proposed framework is demonstrated on two well-known datasets, where we outperform SOTA on WSI classification, gaining a +2.7% AUC and +3.7% accuracy on the popular Camelyon16 benchmark.

The field of multi-object tracking has recently seen a renewed interest in the good old schema of tracking-by-detection, as its simplicity and strong priors spare it from the complex design and painful babysitting of tracking-by-attention approaches. In view of this, we aim at extending tracking-by-detection to multi-modal settings, where a comprehensive cost has to be computed from heterogeneous information e.g., 2D motion cues, visual appearance, and pose estimates. More precisely, we follow a case study where a rough estimate of 3D information is also available and must be merged with other traditional metrics (e.g., the IoU). To achieve that, recent approaches resort to either simple rules or complex heuristics to balance the contribution of each cost. However, i) they require careful tuning of tailored hyperparameters on a hold-out set, and ii) they imply these costs to be independent, which does not hold in reality. We address these issues by building upon an elegant probabilistic formulation, which considers the cost of a candidate association as the negative log-likelihood yielded by a deep density estimator, trained to model the conditional joint probability distribution of correct associations. Our experiments, conducted on both simulated and real benchmarks, show that our approach consistently enhances the performance of several tracking-by-detection algorithms.

We introduce Neuro-Symbolic Continual Learning, where a model has to solve a sequence of neuro-symbolic tasks, that is, it has to map sub-symbolic inputs to high-level concepts and compute predictions by reasoning consistently with prior knowledge. Our key observation is that neuro-symbolic tasks, although different, often share concepts whose semantics remains stable over time. Traditional approaches fall short: existing continual strategies ignore knowledge altogether, while stock neuro-symbolic architectures suffer from catastrophic forgetting. We show that leveraging prior knowledge by combining neuro-symbolic architectures with continual strategies does help avoid catastrophic forgetting, but also that doing so can yield models affected by reasoning shortcuts. These undermine the semantics of the acquired concepts, even when detailed prior knowledge is provided upfront and inference is exact, and in turn continual performance. To overcome these issues, we introduce COOL, a COncept-level cOntinual Learning strategy tailored for neuro-symbolic continual problems that acquires high-quality concepts and remembers them over time. Our experiments on three novel benchmarks highlights how COOL attains sustained high performance on neuro-symbolic continual learning tasks in which other strategies fail.

Denoising Diffusion Probabilistic Models have shown an impressive generation quality although their long sampling chain leads to high computational costs. In this paper, we observe that a long sampling chain also leads to an error accumulation phenomenon, which is similar to the exposure bias problem in autoregressive text generation. Specifically, we note that there is a discrepancy between training and testing, since the former is conditioned on the ground truth samples, while the latter is conditioned on the previously generated results. To alleviate this problem, we propose a very simple but effective training regularization, consisting in perturbing the ground truth samples to simulate the inference time prediction errors. We empirically show that, without affecting the recall and precision, the proposed input perturbation leads to a significant improvement in the sample quality while reducing both the training and the inference times. For instance, on CelebA 64x64, we achieve a new state-of-the-art FID score of 1.27, while saving 37.5% of the training time. The code is available at https://github.com/forever208/DDPM-IP

The staple of human intelligence is the capability of acquiring knowledge in a continuous fashion. In stark contrast, Deep Networks forget catastrophically and, for this reason, the sub-field of Class-Incremental Continual Learning fosters methods that learn a sequence of tasks incrementally, blending sequentially-gained knowledge into a comprehensive prediction. This work aims at assessing and overcoming the pitfalls of our previous proposal Dark Experience Replay (DER), a simple and effective approach that combines rehearsal and Knowledge Distillation. Inspired by the way our minds constantly rewrite past recollections and set expectations for the future, we endow our model with the abilities to i) revise its replay memory to welcome novel information regarding past data ii) pave the way for learning yet unseen classes. We show that the application of these strategies leads to remarkable improvements; indeed, the resulting method - termed eXtended-DER (X-DER) - outperforms the state of the art on both standard benchmarks (such as CIFAR-100 and miniImageNet) and a novel one here introduced. To gain a better understanding, we further provide extensive ablation studies that corroborate and extend the findings of our previous research (e.g., the value of Knowledge Distillation and flatter minima in continual learning setups). We make our results fully reproducible; the codebase is available at https://github.com/aimagelab/mammoth.

The occurrence of West Nile Virus (WNV) represents one of the most common mosquito-borne zoonosis viral infections. Its circulation is usually associated with climatic and environmental conditions suitable for vector proliferation and virus replication. On top of that, several statistical models have been developed to shape and forecast WNV circulation: in particular, the recent massive availability of Earth Observation (EO) data, coupled with the continuous advances in the field of Artificial Intelligence, offer valuable opportunities. In this paper, we seek to predict WNV circulation by feeding Deep Neural Networks (DNNs) with satellite images, which have been extensively shown to hold environmental and climatic features. Notably, while previous approaches analyze each geographical site independently, we propose a spatial-aware approach that considers also the characteristics of close sites. Specifically, we build upon Graph Neural Networks (GNN) to aggregate features from neighbouring places, and further extend these modules to consider multiple relations, such as the difference in temperature and soil moisture between two sites, as well as the geographical distance. Moreover, we inject time-related information directly into the model to take into account the seasonality of virus spread. We design an experimental setting that combines satellite images – from Landsat and Sentinel missions – with ground truth observations of WNV circulation in Italy. We show that our proposed Multi-Adjacency Graph Attention Network (MAGAT) consistently leads to higher performance when paired with an appropriate pre-training stage. Finally, we assess the importance of each component of MAGAT in our ablation studies.

Accurate prediction of future human positions is an essential task for modern video-surveillance systems. Current state-of-the-art models usually rely on a “history” of past tracked locations (e.g., 3 to 5 seconds) to predict a plausible sequence of future locations (e.g., up to the next 5 seconds). We feel that this common schema neglects critical traits of realistic applications: as the collection of input trajectories involves machine perception (i.e., detection and tracking), incorrect detection and fragmentation errors may accumulate in crowded scenes, leading to tracking drifts. On this account, the model would be fed with corrupted and noisy input data, thus fatally affecting its prediction performance. In this regard, we focus on delivering accurate predictions when only few input observations are used, thus potentially lowering the risks associated with automatic perception. To this end, we conceive a novel distillation strategy that allows a knowledge transfer from a teacher network to a student one, the latter fed with fewer observations (just two ones). We show that a properly defined teacher super-vision allows a student network to perform comparably to state-of-the-art approaches that demand more observations. Besides, extensive experiments on common trajectory forecasting datasets highlight that our student network better generalizes to unseen scenarios.

In Continual Learning (CL), a neural network is trained on a stream of data whose distribution changes over time. In this context, the main problem is how to learn new information without forgetting old knowledge (i.e., Catastrophic Forgetting). Most existing CL approaches focus on finding solutions to preserve acquired knowledge, so working on the "past" of the model. However, we argue that as the model has to continually learn new tasks, it is also important to put focus on the "present" knowledge that could improve following tasks learning. In this paper we propose a new, simple, CL algorithm that focuses on solving the current task in a way that might facilitate the learning of the next ones. More specifically, our approach combines the main data stream with a secondary, diverse and uncorrelated stream, from which the network can draw auxiliary knowledge. This helps the model from different perspectives, since auxiliary data may contain useful features for the current and the next tasks and incoming task classes can be mapped onto auxiliary classes. Furthermore, the addition of data to the current task is implicitly making the classifier more robust as we are forcing the extraction of more discriminative features. Our method can outperform existing state-of-the-art models on the most common CL Image Classification benchmarks.

Rehearsal approaches enjoy immense popularity with Continual Learning (CL) practitioners. These methods collect samples from previously encountered data distributions in a small memory buffer; subsequently, they repeatedly optimize on the latter to prevent catastrophic forgetting. This work draws attention to a hidden pitfall of this widespread practice: repeated optimization on a small pool of data inevitably leads to tight and unstable decision boundaries, which are a major hindrance to generalization. To address this issue, we propose Lipschitz-DrivEn Rehearsal (LiDER), a surrogate objective that induces smoothness in the backbone network by constraining its layer-wise Lipschitz constants w.r.t. replay examples. By means of extensive experiments, we show that applying LiDER delivers a stable performance gain to several state-of-the-art rehearsal CL methods across multiple datasets, both in the presence and absence of pre-training. Through additional ablative experiments, we highlight peculiar aspects of buffer overfitting in CL and better characterize the effect produced by LiDER. Code is available at https://github.com/aimagelab/LiDER.

In this article we introduce a new self-supervised, semi-parametric approach for synthesizing novel views of a vehicle starting from a single monocular image. Differently from parametric (i.e., entirely learning-based) methods, we show how a-priori geometric knowledge about the object and the 3D world can be successfully integrated into a deep learning based image generation framework. As this geometric component is not learnt, we call our approach semi-parametric. In particular, we exploit man-made object symmetry and piece-wise planarity to integrate rich a-priori visual information into the novel viewpoint synthesis process. An Image Completion Network (ICN) is then trained to generate a realistic image starting from this geometric guidance. This careful blend between parametric and non-parametric components allows us to i) operate in a real-world scenario, ii) preserve high-frequency visual information such as textures, iii) handle truly arbitrary 3D roto-translations of the input, and iv) perform shape transfer to completely different 3D models. Eventually, we show that our approach can be easily complemented with synthetic data and extended to other rigid objects with completely different topology, even in presence of concave structures and holes (e.g., chairs). A comprehensive experimental analysis against state-of-the-art competitors shows the efficacy of our method both from a quantitative and a perceptive point of view. Supplementary material, animated results, code, and data are available at: https://github.com/ndrplz/semiparametric.

Continual Learning (CL) investigates how to train Deep Networks on a stream of tasks without incurring forgetting. CL settings proposed in literature assume that every incoming example is paired with ground-truth annotations. However, this clashes with many real-world applications: gathering labeled data, which is in itself tedious and expensive, becomes infeasible when data flow as a stream. This work explores Continual Semi-Supervised Learning (CSSL): here, only a small fraction of labeled input examples are shown to the learner. We assess how current CL methods (e.g.: EWC, LwF, iCaRL, ER, GDumb, DER) perform in this novel and challenging scenario, where overfitting entangles forgetting. Subsequently, we design a novel CSSL method that exploits metric learning and consistency regularization to leverage unlabeled examples while learning. We show that our proposal exhibits higher resilience to diminishing supervision and, even more surprisingly, relying only on 25% supervision suffices to outperform SOTA methods trained under full supervision.

Deep learning-based methods for video pedestrian detection and tracking require large volumes of training data to achieve good performance. However, data acquisition in crowded public environments raises data privacy concerns – we are not allowed to simply record and store data without the explicit consent of all participants. Furthermore, the annotation of such data for computer vision applications usually requires a substantial amount of manual effort, especially in the video domain. Labeling instances of pedestrians in highly crowded scenarios can be challenging even for human annotators and may introduce errors in the training data. In this paper, we study how we can advance different aspects of multi-person tracking using solely synthetic data. To this end, we generate MOTSynth, a large, highly diverse synthetic dataset for object detection and tracking using a rendering game engine. Our experiments show that MOTSynth can be used as a replacement for real data on tasks such as pedestrian detection, re-identification, segmentation, and tracking.

Anticipating human motion in crowded scenarios is essential for developing intelligent transportation systems, social-aware robots and advanced video surveillance applications. A key component of this task is represented by the inherently multi-modal nature of human paths which makes socially acceptable multiple futures when human interactions are involved. To this end, we propose a generative architecture for multi-future trajectory predictions based on Conditional Variational Recurrent Neural Networks (C-VRNNs). Conditioning mainly relies on prior belief maps, representing most likely moving directions and forcing the model to consider past observed dynamics in generating future positions. Human interactions are modelled with a graph-based attention mechanism enabling an online attentive hidden state refinement of the recurrent estimation. To corroborate our model, we perform extensive experiments on publicly-available datasets (e.g., ETH/UCY, Stanford Drone Dataset, STATS SportVU NBA, Intersection Drone Dataset and TrajNet++) and demonstrate its effectiveness in crowded scenes compared to several state-of-the-art methods.

Action Detection is a complex task that aims to detect and classify human actions in video clips. Typically, it has been addressed by processing fine-grained features extracted from a video classification backbone. Recently, thanks to the robustness of object and people detectors, a deeper focus has been added on relationship modeling. Following this line, we propose a graph-based framework to learn high-level interactions between people and objects, in both space and time. In our formulation, spatio-temporal relationships are learned through self-attention on a multi-layer graph structure which can connect entities from consecutive clips, thus considering long-range spatial and temporal dependencies. The proposed module is backbone independent by design and does not require end-to-end training. Extensive experiments are conducted on the AVA dataset, where our model demonstrates state-of-the-art results and consistent improvements over baselines built with different backbones. Code is publicly available at https://github.com/aimagelab/STAGE_action_detection.

Convolutional Neural Networks experience catastrophic forgetting when optimized on a sequence of learning problems: as they meet the objective of the current training examples, their performance on previous tasks drops drastically. In this work, we introduce a novel framework to tackle this problem with conditional computation. We equip each convolutional layer with task-specific gating modules, selecting which filters to apply on the given input. This way, we achieve two appealing properties. Firstly, the execution patterns of the gates allow to identify and protect important filters, ensuring no loss in the performance of the model for previously learned tasks. Secondly, by using a sparsity objective, we can promote the selection of a limited set of kernels, allowing to retain sufficient model capacity to digest new tasks. Existing solutions require, at test time, awareness of the task to which each example belongs to. This knowledge, however, may not be available in many practical scenarios. Therefore, we additionally introduce a task classifier that predicts the task label of each example, to deal with settings in which a task oracle is not available. We validate our proposal on four continual learning datasets. Results show that our model consistently outperforms existing methods both in the presence and the absence of a task oracle. Notably, on Split SVHN and Imagenet-50 datasets, our model yields up to 23.98% and 17.42% improvement in accuracy w.r.t. competing methods.

In this paper we present a novel approach for bottom-up multi-person 3D human pose estimation from monocular RGB images. We propose to use high resolution volumetric heatmaps to model joint locations, devising a simple and effective compression method to drastically reduce the size of this representation. At the core of the proposed method lies our Volumetric Heatmap Autoencoder, a fully-convolutional network tasked with the compression of ground-truth heatmaps into a dense intermediate representation. A second model, the Code Predictor, is then trained to predict these codes, which can be decompressed at test time to re-obtain the original representation. Our experimental evaluation shows that our method performs favorably when compared to state of the art on both multi-person and single-person 3D human pose estimation datasets and, thanks to our novel compression strategy, can process full-HD images at the constant runtime of 8 fps regardless of the number of subjects in the scene. Code and models are publicly available.

In Continual Learning, a Neural Network is trained on a stream of data whose distribution shifts over time. Under these assumptions, it is especially challenging to improve on classes appearing later in the stream while remaining accurate on previous ones. This is due to the infamous problem of catastrophic forgetting, which causes a quick performance degradation when the classifier focuses on learning new categories. Recent literature proposed various approaches to tackle this issue, often resorting to very sophisticated techniques. In this work, we show that naive rehearsal can be patched to achieve similar performance. We point out some shortcomings that restrain Experience Replay (ER) and propose five tricks to mitigate them. Experiments show that ER, thus enhanced, displays an accuracy gain of 51.2 and 26.9 percentage points on the CIFAR-10 and CIFAR-100 datasets respectively (memory buffer size 1000). As a result, it surpasses current state-of-the-art rehearsal-based methods.

The ability to detect, localize and classify objects that are anomalies is a challenging task in the computer vision community. In this paper, we tackle these tasks developing a framework to automatically inspect the railway during the night. Specifically, it is able to predict the presence, the image coordinates and the class of obstacles. To deal with the low-light environment, the framework is based on thermal images and consists of three different modules that address the problem of detecting anomalies, predicting their image coordinates and classifying them. Moreover, due to the absolute lack of publicly-released datasets collected in the railway context for anomaly detection, we introduce a new multi-modal dataset, acquired from a rail drone, used to evaluate the proposed framework. Experimental results confirm the accuracy of the framework and its suitability, in terms of computational load, performance, and inference time, to be implemented on a self-powered inspection system.

Understanding human motion behaviour is a critical task for several possible applications like self-driving cars or social robots, and in general for all those settings where an autonomous agent has to navigate inside a human-centric environment. This is non-trivial because human motion is inherently multi-modal: given a history of human motion paths, there are many plausible ways by which people could move in the future. Additionally, people activities are often driven by goals, e.g. reaching particular locations or interacting with the environment. We address the aforementioned aspects by proposing a new recurrent generative model that considers both single agents' future goals and interactions between different agents. The model exploits a double attention-based graph neural network to collect information about the mutual influences among different agents and to integrate it with data about agents' possible future objectives. Our proposal is general enough to be applied to different scenarios: the model achieves state-of-the-art results in both urban environments and also in sports applications.

In this work we propose a deep learning pipeline to predict the visual future appearance of an urban scene. Despite recent advances, generating the entire scene in an end-to-end fashion is still far from being achieved. Instead, here we follow a two stages approach, where interpretable information is included in the loop and each actor is modelled independently. We leverage a per-object novel view synthesis paradigm; i.e. generating a synthetic representation of an object undergoing a geometrical roto-translation in the 3D space. Our model can be easily conditioned with constraints (e.g. input trajectories) provided by state-of-the-art tracking methods or by the user itself. This allows us to generate a set of diverse realistic futures starting from the same input in a multi-modal fashion. We visually and quantitatively show the superiority of this approach over traditional end-to-end scene-generation methods on CityFlow, a challenging real world dataset.

The recently proposed action spotting task consists in finding the exact timestamp in which an event occurs. This task fits particularly well for soccer videos, where events correspond to salient actions strictly defined by soccer rules (a goal occurs when the ball crosses the goal line). In this paper, we devise a lightweight and modular network for action spotting, which can simultaneously predict the event label and its temporal offset using the same underlying features. We enrich our model with two training strategies: the first one for data balancing and uniform sampling, the second for masking ambiguous frames and keeping the most discriminative visual cues. When tested on the SoccerNet dataset and using standard features, our full proposal exceeds the current state of the art by 3 Average-mAP points. Additionally, it reaches a gain of more than 10 Average-mAP points on the test set when fine-tuned in combination with a strong 2D backbone.

The recent growth in the number of satellite images fosters the development of effective deep-learning techniques for Remote Sensing (RS). However, their full potential is untapped due to the lack of large annotated datasets. Such a problem is usually countered by fine-tuning a feature extractor that is previously trained on the ImageNet dataset. Unfortunately, the domain of natural images differs from the RS one, which hinders the final performance. In this work, we propose to learn meaningful representations from satellite imagery, leveraging its high-dimensionality spectral bands to reconstruct the visible colors. We conduct experiments on land cover classification (BigEarthNet) and West Nile Virus detection, showing that colorization is a solid pretext task for training a feature extractor. Furthermore, we qualitatively observe that guesses based on natural images and colorization rely on different parts of the input. This paves the way to an ensemble model that eventually outperforms both the above-mentioned techniques.

Continual Learning has inspired a plethora of approaches and evaluation settings; however, the majority of them overlooks the properties of a practical scenario, where the data stream cannot be shaped as a sequence of tasks and offline training is not viable. We work towards General Continual Learning (GCL), where task boundaries blur and the domain and class distributions shift either gradually or suddenly. We address it through mixing rehearsal with knowledge distillation and regularization; our simple baseline, Dark Experience Replay, matches the network's logits sampled throughout the optimization trajectory, thus promoting consistency with its past. By conducting an extensive analysis on both standard benchmarks and a novel GCL evaluation setting (MNIST-360), we show that such a seemingly simple baseline outperforms consolidated approaches and leverages limited resources. We further explore the generalization capabilities of our objective, showing its regularization being beneficial beyond mere performance.

Depth cameras allow to set up reliable solutions for people monitoring and behavior understanding, especially when unstable or poor illumination conditions make unusable common RGB sensors. Therefore, we propose a complete framework for the estimation of the head and shoulder pose based on depth images only. A head detection and localization module is also included, in order to develop a complete end-to-end system. The core element of the framework is a Convolutional Neural Network, called POSEidon+, that receives as input three types of images and provides the 3D angles of the pose as output. Moreover, a Face-from-Depth component based on a Deterministic Conditional GAN model is able to hallucinate a face from the corresponding depth image. We empirically demonstrate that this positively impacts the system performances. We test the proposed framework on two public datasets, namely Biwi Kinect Head Pose and ICT-3DHP, and on Pandora, a new challenging dataset mainly inspired by the automotive setup. Experimental results show that our method overcomes several recent state-of-art works based on both intensity and depth input data, running in real-time at more than 30 frames per second.

We present a novel and hierarchical approach for supervised classification of signals spanning over a fixed graph, reflecting shared properties of the dataset. To this end, we introduce a Convolutional Cluster Pooling layer exploiting a multi-scale clustering in order to highlight, at different resolutions, locally connected regions on the input graph. Our proposal generalises well-established neural models such as Convolutional Neural Networks (CNNs) on irregular and complex domains, by means of the exploitation of the weight sharing property in a graph-oriented architecture. In this work, such property is based on the centrality of each vertex within its soft-assigned cluster. Extensive experiments on NTU RGB+D, CIFAR-10 and 20NEWS demonstrate the effectiveness of the proposed technique in capturing both local and global patterns in graph-structured data out of different domains.

Cloud computing data centers are growing in size and complexity to the point where monitoring and management of the infrastructure become a challenge due to scalability issues. A possible approach to cope with the size of such data centers is to identify VMs exhibiting a similar behavior. Existing literature demonstrated that clustering together VMs that show a similar behavior may improve the scalability of both monitoring andmanagement of a data center. However, available techniques suffer from a trade-off between accuracy and time to achieve this result. Throughout this paper we propose a different approach where, instead of an unsupervised clustering, we rely on classifiers based on deep learning techniques to assigna newly deployed VMs to a cluster of already-known VMs. The two proposed classifiers, namely DeepConv and DeepFFT use a convolution neural network and (in the latter model) exploits Fast Fourier Transformation to classify the VMs. Our proposal is validated using a set of traces describing the behavior of VMs from a realcloud data center. The experiments compare our proposal with state-of-the-art solutions available in literature, demonstrating that our proposal achieve better performance. Furthermore, we show that our solution issignificantly faster than the alternatives as it can produce a perfect classification even with just a few samples of data, making our proposal viable also toclassify on-demand VMs that are characterized by a short life span.

Novelty detection is commonly referred as the discrimination of observations that do not conform to a learned model of regularity. Despite its importance in different application settings, designing a novelty detector is utterly complex due to the unpredictable nature of novelties and its inaccessibility during the training procedure, factors which expose the unsupervised nature of the problem. In our proposal, we design a general unsupervised framework where we equip a deep autoencoder with a parametric density estimator that learns the probability distribution underlying the latent representations with an autoregressive procedure. We show that a maximum likelihood objective, optimized in conjunction with the reconstruction of normal samples, effectively acts as a regularizer for the task at hand, by minimizing the differential entropy of the distribution spanned by latent vectors. In addition to providing a very general formulation, extensive experiments of our model on publicly available datasets deliver on-par or superior performances if compared to state-of-the-art methods in one-class and in video anomaly detection settings. Differently from our competitors, we remark that our proposal does not make any assumption about the nature of the novelties, making our work easily applicable to disparate contexts.

When you see a person in a crowd, occluded by other persons, you miss visual information that can be used to recognize, re-identify or simply classify him or her. You can imagine its appearance given your experience, nothing more. Similarly, AI solutions can try to hallucinate missing information with specific deep learning architectures, suitably trained with people with and without occlusions. The goal of this work is to generate a complete image of a person, given an occluded version in input, that should be a) without occlusion b) similar at pixel level to a completely visible people shape c) capable to conserve similar visual attributes (e.g. male/female) of the original one. For the purpose, we propose a new approach by integrating the state-of-the-art of neural network architectures, namely U-nets and GANs, as well as discriminative attribute classification nets, with an architecture specifically designed to de-occlude people shapes. The network is trained to optimize a Loss function which could take into account the aforementioned objectives. As well we propose two datasets for testing our solution: the first one, occluded RAP, created automatically by occluding real shapes of the RAP dataset created by Li et al. (2016) (which collects also attributes of the people aspect); the second is a large synthetic dataset, AiC, generated in computer graphics with data extracted from the GTA video game, that contains 3D data of occluded objects by construction. Results are impressive and outperform any other previous proposal. This result could be an initial step to many further researches to recognize people and their behavior in an open crowded world.

In this work we aim to predict the driver's focus of attention. The goal is to estimate what a person would pay attention to while driving, and which part of the scene around the vehicle is more critical for the task. To this end we propose a new computer vision model based on a multi-branch deep architecture that integrates three sources of information: raw video, motion and scene semantics. We also introduce DR(eye)VE, the largest dataset of driving scenes for which eye-tracking annotations are available. This dataset features more than 500,000 registered frames, matching ego-centric views (from glasses worn by drivers) and car-centric views (from roof-mounted camera), further enriched by other sensors measurements. Results highlight that several attention patterns are shared across drivers and can be reproduced to some extent. The indication of which elements in the scene are likely to capture the driver's attention may benefit several applications in the context of human-vehicle interaction and driver attention analysis.

Dense optical flow estimation is complex and time consuming, with state-of-the-art methods relying either on large synthetic data sets or on pipelines requiring up to a few minutes per frame pair. In this paper, we address the problem of optical flow estimation in the automotive scenario in a self-supervised manner. We argue that optical flow can be cast as a geometrical warping between two successive video frames and devise a deep architecture to estimate such transformation in two stages. First, a dense pixel-level flow is computed with a projective bootstrap on rigid surfaces. We show how such global transformation can be approximated with a homography and extend spatial transformer layers so that they can be employed to compute the flow field implied by such transformation. Subsequently, we refine the prediction by feeding a second, deeper network that accounts for moving objects. A final reconstruction loss compares the warping of frame Xtwith the subsequent frame Xt+1and guides both estimates. The model has the speed advantages of end-to-end deep architectures while achieving competitive performances, both outperforming recent unsupervised methods and showing good generalization capabilities on new automotive data sets.

Can faces acquired by low-cost depth sensors be useful to catch some characteristic details of the face? Typically the answer is no. However, new deep architectures can generate RGB images from data acquired in a different modality, such as depth data. In this paper, we propose a new Deterministic Conditional GAN, trained on annotated RGB-D face datasets, effective for a face-to-face translation from depth to RGB. Although the network cannot reconstruct the exact somatic features for unknown individual faces, it is capable to reconstruct plausible faces; their appearance is accurate enough to be used in many pattern recognition tasks. In fact, we test the network capability to hallucinate with some Perceptual Probes, as for instance face aspect classification or landmark detection. Depth face can be used in spite of the correspondent RGB images, that often are not available due to difficult luminance conditions. Experimental results are very promising and are as far as better than previously proposed approaches: this domain translation can constitute a new way to exploit depth data in new future applications.

In this paper we propose a deep architecture for detecting people attributes (e.g. gender, race, clothing ...) in surveillance contexts. Our proposal explicitly deal with poor resolution and occlusion issues that often occur in surveillance footages by enhancing the images by means of Deep Convolutional Generative Adversarial Networks (DCGAN). Experiments show that by combining both our Generative Reconstruction and Deep Attribute Classification Network we can effectively extract attributes even when resolution is poor and in presence of strong occlusions up to 80% of the whole person figure.

Despite the advent of autonomous cars, it's likely - at least in the near future - that human attention will still maintain a central role as a guarantee in terms of legal responsibility during the driving task. In this paper we study the dynamics of the driver's gaze and use it as a proxy to understand related attentional mechanisms. First, we build our analysis upon two questions: where and what the driver is looking at? Second, we model the driver's gaze by training a coarse-to-fine convolutional network on short sequences extracted from the DR(eye)VE dataset. Experimental comparison against different baselines reveal that the driver's gaze can indeed be learnt to some extent, despite (i) being highly subjective and (ii) having only one driver's gaze available for each sequence due to the irreproducibility of the scene. Eventually, we advocate for a new assisted driving paradigm which suggests to the driver, with no intervention, where she should focus her attention.

We propose a method for tracking groups from single and multiple cameras with disjointed fields of view. Our formulation follows the tracking-by-detection paradigm in which groups are the atomic entities and are linked over time to form long and consistent trajectories. To this end, we formulate the problem as a supervised clustering problem in which a structural SVM classifier learns a similarity measure appropriate for group entities. Multicamera group tracking is handled inside the framework by adopting an orthogonal feature encoding that allows the classifier to learn inter- and intra-camera feature weights differently. Experiments were carried out on a novel annotated group tracking data set, the DukeMTMC-Groups data set. Since this is the first data set on the problem, it comes with the proposal of a suitable evaluation measure. Results of adopting learning for the task are encouraging, scoring a +15% improvement in F1measure over a nonlearning-based clustering baseline. To the best of our knowledge, this is the first proposal of its kind dealing with multicamera group tracking.

Despite prejudice cannot be directly observed, nonverbal behaviours provide profound hints on people inclinations. In this paper, we use recent sensing technologies and machine learning techniques to automatically infer the results of psychological questionnaires frequently used to assess implicit prejudice. In particular, we recorded 32 students discussing with both white and black collaborators. Then, we identified a set of features allowing automatic extraction and measured their degree of correlation with psychological scores. Results confirmed that automated analysis of nonverbal behaviour is actually possible thus paving the way for innovative clinical tools and eventually more secure societies.

Modern crowd theories agree that collective behavior is the result of the underlying interactions among small groups of individuals. In this work, we propose a novel algorithm for detecting social groups in crowds by means of a Correlation Clustering procedure on people trajectories. The affinity between crowd members is learned through an online formulation of the Structural SVM framework and a set of specifically designed features characterizing both their physical and social identity, inspired by Proxemic theory, Granger causality, DTW and Heat-maps. To adhere to sociological observations, we introduce a loss function ( G -MITRE) able to deal with the complexity of evaluating group detection performances. We show our algorithm achieves state-of-the-art results when relying on both ground truth trajectories and tracklets previously extracted by available detector/tracker systems.

Long-term tracking of people in unconstrained scenarios is still an open problem due to the absence of constant elements in the problem setting. The camera, when active, may move and the appearance of both the background and the target may change abruptly, leading to the inadequacy of most standard tracking techniques. We propose to exploit a learning approach that considers the tracking task as a semisupervised learning problem. Given few target samples, the aim is to search for the target occurrences in the video stream, reinterpreting the problem as label propagation on a similarity graph. We propose a solution based on graph transduction that iteratively works frame by frame. In addition, to avoid drifting, we introduce an update strategy based on an evolutionary clustering technique that chooses the visual templates that better describe target appearance, evolving the model during the processing of the video. Since we model people's appearance by means of covariance matrices on color and gradient information, our framework is directly related to structure learning on Riemannian manifolds. Tests on publicly available data sets and comparisons with state-of-the-art techniques allow us to conclude that our solution exhibits interesting performances in terms of tracking precision and recall in most of the considered scenarios.

Conventional experiments on MTT are built upon the belief that fixing the detections to different trackers is sufficient to obtain a fair comparison. In this work we argue how the true behavior of a tracker is exposed when evaluated by varying the input detections rather than by fixing them. We propose a systematic and reproducible protocol and a MATLAB toolbox for generating synthetic data starting from ground truth detections, a proper set of metrics to understand and compare trackers peculiarities and respective visualization solutions.

Online Multiple Target Tracking (MTT) is often addressed within the tracking-by-detection paradigm. Detections are previously extracted independently in each frame and then objects trajectories are built by maximizing specifically designed coherence functions. Nevertheless, ambiguities arise in presence of occlusions or detection errors. In this paper we claim that the ambiguities in tracking could be solved by a selective use of the features, by working with more reliable features if possible and exploiting a deeper representation of the target only if necessary. To this end, we propose an online divide and conquer tracker for static camera scenes, which partitions the assignment problem in local subproblems and solves them by selectively choosing and combining the best features. The complete framework is cast as a structural learning task that unifies these phases and learns tracker parameters from examples. Experiments on two different datasets highlights a significant improvement of tracking performances (MOTA +10%) over the state of the art.

In this paper we present a new method for head pose real-time estimation in ego-vision scenarios that is a key step in the understanding of social interactions. In order to robustly detect head under changing aspect ratio, scale and orientation we use and extend the Hough-Based Tracker which allows to follow simultaneously each subject in the scene. In an ego-vision scenario where a group interacts in a discussion, each subject's head orientation will be more likely to remain focused for a while on the person who has the floor. In order to encode this behavior we include a stateful Hidden Markov Model technique that enforces the predicted pose with the temporal coherence from a video sequence. We extensively test our approach on several indoor and outdoor ego-vision videos with high illumination variations showing its validity and outperforming other recent related state of the art approaches.

Despite pattern recognition methods for human behavioral analysis has flourished in the last decade, animal behavioral analysis has been almost neglected. Those few approaches are mostly focused on preserving livestock economic value while attention on the welfare of companion animals, like dogs, is now emerging as a social need. In this work, following the analogy with human behavior recognition, we propose a system for recognizing body parts of dogs kept in pens. We decide to adopt both 2D and 3D features in order to obtain a rich description of the dog model. Images are acquired using the Microsoft Kinect to capture the depth map images of the dog. Upon depth maps a Structural Support Vector Machine (SSVM) is employed to identify the body parts using both 3D features and 2D images. The proposal relies on a kernelized discriminative structural classificator specifically tailored for dogs independently from the size and breed. The classification is performed in an online fashion using the LaRank optimization technique to obtaining real time performances. Promising results have emerged during the experimental evaluation carried out at a dog shelter, managed by IZSAM, in Teramo, Italy.

There is a large variety of trackers, which have been proposed in the literature during the last two decades with some mixed success. Object tracking in realistic scenarios is a difficult problem, therefore, it remains a most active area of research in computer vision. A good tracker should perform well in a large number of videos involving illumination changes, occlusion, clutter, camera motion, low contrast, specularities, and at least six more aspects. However, the performance of proposed trackers have been evaluated typically on less than ten videos, or on the special purpose datasets. In this paper, we aim to evaluate trackers systematically and experimentally on 315 video fragments covering above aspects. We selected a set of nineteen trackers to include a wide variety of algorithms often cited in literature, supplemented with trackers appearing in 2010 and 2011 for which the code was publicly available. We demonstrate that trackers can be evaluated objectively by survival curves, Kaplan Meier statistics, and Grubs testing. We find that in the evaluation practice the F-score is as effective as the object tracking accuracy (OTA) score. The analysis under a large variety of circumstances provides objective insight into the strengths and weaknesses of trackers.

Group detection in crowds will play a key role in future behavior analysis surveillance systems. In this work we build a new Structural SVM-based learning framework able to solve the group detection task by exploiting annotated video data to deduce a sociologically motivated distance measure founded on Hall's proxemics and Granger's causality. We improve over state-of-the-art results even in the most crowded test scenarios, while keeping the classification time affordable for quasi-real time applications. A new scoring scheme specifically designed for the group detection task is also proposed.

We propose a formulation of people tracking problem as a Transductive Learning (TL) problem. TL is an effective semi-supervised learning technique by which many classification problems have been recently reinterpreted as learning labels from incomplete datasets. In our proposal the joint exploitation of spectral graph theory and Riemannian manifold learning tools leads to the formulation of a robust approach for appearance based tracking in Video Surveillance scenarios. The key advantage of the presented method is a continuously updated model of the tracked target, used in the TL process, that allows to on-line learn the target visual appearance and consequently to improve the tracker accuracy. Experiments on public datasets show an encouraging advancement over alternative state-of the-art techniques.

People trajectory analysis is a recurrent task in many pattern recognition applications, such as surveillance, behavior analysis, video annotation, and many others. In this paper, we propose a new framework for analyzing trajectory shape, invariant to spatial shifts of the people motion in the scene. In order to cope with the noise and the uncertainty of the trajectory samples, we propose to describe the trajectories as a sequence of angles modeled by distributions of circular statistics, i.e., a mixture of von Mises (MovM) distributions. To deal with MovM, we define a new specific expectation-maximization (EM) algorithm for estimating the parameters and derive a closed form of the Bhattacharyya distance between single von Mises pdfs. Trajectories are then modeled with a sequence of symbols, corresponding to the most suitable distribution in the mixture, and compared each other after a global alignment procedure to cope with trajectories of different lengths. The trajectories in the training set are clustered according to their shape similarity in an off-line phase, and testing trajectories are then classified with a specific on-line EM, based on sufficient statistics. The approach is particularly suitable for classifying people trajectories in video surveillance, searching for abnormal (i.e., infrequent) paths. Tests on synthetic and real data are provided with also a complete comparison with other circular statistical and alignment methods.

This paper presents a method for comparing people trajectories for video surveillance applications, based on semi-directional statistics. In fact, the modelling of a trajectory as a sequence of angles, speeds and time lags, requires the use of a statistical tool capable to jointly consider periodic and linear variables. Our statistical method is compared with two state-of-the-art methods.

This paper proposes a system for people trajectory shape analysis by exploiting a statistical approach which accounts for sequences of both directional (the directions of the trajectory) and linear (the speeds) data. A semi-directional distribution (AWLG - Approximated Wrapped and Linear Gaussian) is used with a mixture to find main directions and speeds. A variational version of the mutual information criterion is proposed to prove the statistical dependency of the data. Then, in order to compare data sequences, we define an inexact method with a Kullback-Leibler-based distance measure and employ a global alignment technique is to handle sequences of different lengths and with local shifts or deformations. A comprehensive analysis of variable dependency and parameter estimation techniques are reported and evaluated on both synthetic and real data sets.

Recognizing different actions with a unique approach can be a difficult task. This paper proposes a novel holistic representation of actions that we called "action signature". This 1D trajectory is obtained by parsing the 2D image containing the orientations of the gradient calculated on the motion feature map called motion-history image. In this way, the trajectory is a sketch representation of how the object motion varies in time. A robust statistical framework based on mixtures of von Mises distributions and dynamic programming for sequence alignment are used to compare and classify actions/trajectories. The experimental results show a rather high accuracy in distinguishing quite complicated actions, such as drinking, jumping, or abandoning an object.

The analysis of patterns of movement is a crucial task for several surveillance applications, for instance to classify normal or abnormal people trajectories on the basis of their occurrence. This paper proposes to model the shape of a single trajectory as a sequence of angles described using a mixture of Von Mises (MoVM) distribution. A complete EM (expectation maximization) algorithm is derived for MoVM parameters estimation and an on-line version proposed to meet real time requirement. Maximum-A-Posteriori is used to encode the trajectory as a sequence of symbols corresponding to the MoVM components. Iterative k-medoids clustering groups trajectories in a variable number of similarity classes. The similarity is computed aligning (with dynamic programming) two sequences and considering as symbol-to-symbol distance the Bhattacharyya distance between von Mises distributions. Extensive experiments have been performed on both synthetic and real data.

Smoke detection calls for a reliable and fast distinction between background, moving objects and variable shapes that are recognizable as smoke. In our system we propose a stable background suppression module joined with a smoke detection module working on segmented objects. It exploits two features: the energy variation in wavelet model and a color model of the smoke. The decrease of energy ratio in wavelet domain between background and current image is a clue to detect smoke representing the variations of texture level. A mixture of Gaussians models this texture ratio for temporal evolution. The color model is used as reference to measure the deviation of the current pixel color from the model. The two features have been combined using a Bayesian classifier to detect smoke in the scene. Experiments on real data and a comparison between our background model and Gaussian mixture (MOG) model for smoke detection are presented.

This paper presents a novel and robust approach to consistent labeling for people surveillance in multi-camera systems. A general framework scalable to any number of cameras with overlapped views is devised. An off-line training process automatically computes ground-plane homography and recovers epipolar geometry. When a new object is detected in any one camera, hypotheses for potential matching objects in the other cameras are established. Each of the hypotheses is evaluated using a prior and likelihood value. The prior accounts for the positions of the potential matching objects, while the likelihood is computed by warping the vertical axis of the new object on the field of view of the other cameras and measuring the amount of match. In the likelihood, two contributions (forward and backward) are considered so as to correctly handle the case of groups of people merged into single objects. Eventually, a maximum-a-posteriori approach estimates the best label assignment for the new object. Comparisons with other methods based on homography and extensive outdoor experiments demonstrate that the proposed approach is accurate and robust in coping with segmentation errors and in disambiguating groups.

This paper proposes the use of a mixture of Von Mises distributions to detect abnormal behaviors of moving people. The mixture is created from an unsupervised training set by exploiting k-medoids clustering algorithm based on Bhattacharyya distance between distributions. The extracted medoids are used as modes in the multi-modal mixture whose weights are the priors of the specific medoid. Given the mixture model a new trajectory is verified on the model by considering each direction composing it as independent. Experiments over a real scenario composed of multiple, partially-overlapped cameras are reported.

Logging information on moving objects is crucial in video surveillance systems. Distributed multi-camera systems can provide the appearance of objects/people from different viewpoints and at different resolutions, allowing a more complete and precise logging of the information. This is achieved through consistent labeling to correlate collected information of the same person. This paper proposes a novel approach to consistent labeling also capable to fully characterize groups of people and to manage miss segmentations. The ground-plane homography and the epipolar geometry are automatically learned and exploited to warp objects' principal axes between overlapped cameras. A MAP estimator that exploits two contributions (forward and backward) is used to choose the most probable label configuration to be assigned at the handoff of a new object. Extensive experiments demonstrate the accuracy of the proposed method in detecting single and simultaneous handoffs, miss segmentations, and groups.

Efficient solution to people tracking in distributed video surveillance is requested to monitor crowded and large environments. This paper proposes a novel use of the entry edges of field of view (E/sup 2/oFoV) to solve the consistent labeling problem between partially overlapped views. An automatic and reliable procedure allows obtaining the homographic transformation between two overlapped views, without any manual calibration of the cameras. Through the homography, the consistent labeling is established each time a new track is detected in one of the cameras. A camera transition graph (CTG) is defined to speed up the establishment process by reducing the search space. Experimental results prove the effectiveness of the proposed solution also in challenging conditions.