Federated learning (FL) enables collaborative model training with privacy preservation. Data heterogeneity across edge devices (clients) can cause models to converge to sharp minima, negatively impacting generalization and robustness. Recent approaches use client-side sharpness-aware minimization (SAM) to encourage flatter minima, but the discrepancy between local and global loss landscapes often undermines their effectiveness, as optimizing for local sharpness does not ensure global flatness. This work introduces FedGloSS (Federated Global Server-side Sharpness), a novel FL approach that prioritizes the optimization of global sharpness on the server, using SAM. To reduce communication overhead, FedGloSS cleverly approximates sharpness using the previous global gradient, eliminating the need for additional client communication. Our extensive evaluations demonstrate that FedGloSS consistently achieves flatter minima and better performance compared to state-of-the-art FL methods in various federated vision benchmarks. Code available at github.com/pietrocagnasso/fedgloss.

Federated Learning (FL) emerged as a decentralized paradigm to train models while preserving privacy. However, conventional FL struggles with data heterogeneity and class imbalance, which degrade model performance. Clustered FL balances personalization and decentralized training by grouping clients with analogous data distributions, enabling improved accuracy while adhering to privacy constraints. This approach effectively mitigates the adverse impact of heterogeneity in FL. In this work, we propose a novel clustering method for FL, FedGWC (Federated Gaussian Weighting Clustering), which groups clients based on their data distribution, allowing training of a more robust and personalized model on the identified clusters. FedGWC identifies homogeneous clusters by transforming individual empirical losses to model client interactions with a Gaussian reward mechanism. Additionally, we introduce the Wasserstein Adjusted Score, a new clustering metric for FL to evaluate cluster cohesion with respect to the individual class distribution. Our experiments on benchmark datasets show that FedGWC outperforms existing FL algorithms in cluster quality and classification accuracy, validating the efficacy of our approach.

Recognizing human activities from visual inputs, particularly through a first-person viewpoint, is essential for enabling robots to replicate human behavior. Egocentric vision, characterized by cameras worn by observers, captures diverse changes in illumination, viewpoint, and environment. This variability leads to a notable drop in the performance of Egocentric Action Recognition models when tested in environments not seen during training. In this paper, we tackle these challenges by proposing a domain generalization approach for Egocentric Action Recognition. Our insight is that action sequences often reflect consistent user intent across visual domains. By leveraging action sequences , we aim to enhance the model’s generalization ability across unseen environments. Our proposed method, named SeqDG, introduces a visual-text sequence reconstruction objective (SeqRec) that uses contextual cues from both text and visual inputs to reconstruct the central action of the sequence. Additionally, we enhance the model’s robustness by training it on mixed sequences of actions from different domains (SeqMix). We validate SeqDG on the EGTEA and EPIC-KITCHENS-100 datasets. Results on EPIC-KITCHENS-100, show that SeqDG leads to +2.4% relative average improvement in cross-domain action recognition in unseen environments, and on EGTEA the model achieved +0.6% Top-1 accuracy over SOTA in intra-domain action recognition.

Human activities exhibit a strong correlation between actions and the places where these are performed, such as washing something at a sink. More specifically, in daily living environments we may identify particular locations, hereinafter named activity-centric zones , which may afford a set of homogeneous actions. Their knowledge can serve as a prior to favor vision models to recognize human activities. However, the appearance of these zones is scene-specific, limiting the transferability of this prior information to unfamiliar areas and domains. This problem is particularly relevant in egocentric vision, where the environment takes up most of the image, making it even more difficult to separate the action from the context. In this paper, we discuss the importance of decoupling the domain-specific appearance of activity-centric zones from their universal, domain-agnostic representations, and show how the latter can improve the cross-domain transferability of Egocentric Action Recognition (EAR) models. We validate our solution on the EPIC-Kitchens-100 and Argo1M datasets. • We shed light on the side-effects of co-occurrence biases in egocentric vision. • We present EgoZAR, an action recognition model leveraging environmental affordances. • We achieve SOTA performance on EPIC-Kitchens-100 in the domain generalization setting.

Astronaut photography, spanning six decades of human spaceflight, presents a unique Earth observations dataset with immense value for both scientific research and disaster response. Despite their significance, accurately localizing the geographical extent of these images, which is crucial for effective utilization, poses substantial challenges. Cur-rent, manual localization efforts are time-consuming, mo-tivating the need for automated solutions. We propose a novel approach - leveraging image retrieval - to address this challenge efficiently. We introduce innovative training techniques which contribute to the development of a high-performance model, EarthLoc. We develop six evaluation datasets and perform a comprehensive benchmark comparing EarthLoc to existing methods, showcasing its superior efficiency and accuracy. Our approach marks a signifi-cant advancement in automating the localization of astro-naut photography, which will help bridge a critical gap in Earth observations data. Code and datasets are available at https://github.com/gmberton/EarthLoc.

Pose refinement is an interesting and practically rele-vant research direction. Pose refinement can be used to (1) obtain a more accurate pose estimate from an initial prior (e.g., from retrieval), (2) as pre-processing, i.e., to provide a better starting point to a more expensive pose estimator, (3) as post-processing of a more accurate local-izer. Existing approaches focus on learning features / scene representations for the pose refinement task. This involves training an implicit scene representation or learning features while optimizing a camera pose-based loss. A natural question is whether training specific features / representations is truly necessary or whether similar results can be al-ready achieved with more generic features. In this work, we present a simple approach that combines pre-trained features with a particle filter and a renderable representation of the scene. Despite its simplicity, it achieves state-of-the-art results, demonstrating that one can easily build a pose refiner without the need for specific training. The code is at https://github.com/gali130/mcloc_poseref

Federated Learning (FL) methods often struggle in highly statistically heterogeneous settings. Indeed, non-IID data distributions cause client drift and biased local solutions, particularly pronounced in the final classification layer, negatively impacting convergence speed and accuracy. To address this issue, we introduce *Federated Recursive Ridge Regression* (Fed3R). Our method fits a Ridge Regression classifier computed in closed form leveraging pre-trained features. Fed3R is immune to statistical heterogeneity and is invariant to the sampling order of the clients. Therefore, it proves particularly effective in cross-device scenarios. Furthermore, it is fast and efficient in terms of communication and computation costs, requiring up to two orders of magnitude fewer resources than the competitors. Finally, we propose to leverage the Fed3R parameters as an initialization for a softmax classifier and subsequently fine-tune the model using any FL algorithm (Fed3R with Fine-Tuning, Fed3R+FT). Our findings also indicate that maintaining a fixed classifier aids in stabilizing the training and learning more discriminative features in cross-device settings. Official website: https://fed-3r.github.io/.

Abstract Multi-modal learning has gained significant attention due to its ability to enhance machine learning algorithms. However, it brings challenges related to modality heterogeneity and domain shift. In this work, we address these challenges by proposing a new approach called Relative Norm Alignment (RNA) loss. RNA loss exploits the observation that variations in marginal distributions between modalities manifest as discrepancies in their mean feature norms, and rebalances feature norms across domains, modalities, and classes. This rebalancing improves the accuracy of models on test data from unseen (“target”) distributions. In the context of Unsupervised Domain Adaptation (UDA), we use unlabeled target data to enhance feature transferability. We achieve this by combining RNA loss with an adversarial domain loss and an Information Maximization term that regularizes predictions on target data. We present a comprehensive analysis and ablation of our method for both Domain Generalization and UDA settings, testing our approach on different modalities for tasks such as first and third person action recognition, object recognition, and fatigue detection. Experimental results show that our approach achieves competitive or state-of-the-art performance on the proposed benchmarks, showing the versatility and effectiveness of our method in a wide range of applications.

Segmenting unknown or anomalous object instances is a critical task in autonomous driving applications, and it is approached traditionally as a per-pixel classification problem. However, reasoning individually about each pixel without considering their contextual semantics results in high uncertainty around the objects' boundaries and numerous false positives. We propose a paradigm change by shifting from a per-pixel classification to a mask classification. Our mask-based method, Mask2Anomaly, demonstrates the feasibility of integrating a mask-classification architecture to jointly address anomaly segmentation, open-set semantic segmentation, and open-set panoptic segmentation. Mask2Anomaly includes several technical novelties that are designed to improve the detection of anomalies/unknown objects: i) a global masked attention module to focus individually on the foreground and background regions; ii) a mask contrastive learning that maximizes the margin between an anomaly and known classes; iii) a mask refinement solution to reduce false positives; and iv) a novel approach to mine unknown instances based on the mask- architecture properties. By comprehensive qualitative and qualitative evaluation, we show Mask2Anomaly achieves new state-of-the-art results across the benchmarks of anomaly segmentation, open-set semantic segmentation, and open-set panoptic segmentation.

Visual Place Recognition is a task that aims to predict the place of an image (called query) based solely on its visual features. This is typically done through image retrieval, where the query is matched to the most similar images from a large database of geotagged photos, using learned global descriptors. A major challenge in this task is recognizing places seen from different viewpoints. To overcome this limitation, we propose a new method, called EigenPlaces, to train our neural network on images from different point of views, which embeds viewpoint robustness into the learned global descriptors. The underlying idea is to cluster the training data so as to explicitly present the model with different views of the same points of interest. The selection of this points of interest is done without the need for extra supervision. We then present experiments on the most comprehensive set of datasets in literature, finding that EigenPlaces is able to outperform previous state of the art on the majority of datasets, while requiring 60% less GPU memory for training and using 50% smaller descriptors. The code and trained models for EigenPlaces are available at https://github.com/gmberton/EigenPlaces, while results with any other baseline can be computed with the codebase at https://github.com/gmberton/auto_VPR.

We propose and address a new generalisation problem: can a model trained for action recognition successfully classify actions when they are performed within a previously unseen scenario and in a previously unseen location? To answer this question, we introduce the Action Recognition Generalisation Over scenarios and locations dataset (ARGO1M), which contains 1.1M video clips from the large-scale Ego4D dataset, across 10 scenarios and 13 locations. We demonstrate recognition models struggle to generalise over 10 proposed test splits, each of an unseen scenario in an unseen location. We thus propose CIR, a method to represent each video as a Cross-Instance Reconstruction of videos from other domains. Reconstructions are paired with text narrations to guide the learning of a domain generalisable representation. We provide extensive analysis and ablations on ARGO1M that show CIR outperforms prior domain generalisation works on all test splits. Code and data: https://chiaraplizz.github.io/what-can-a-cook/.

Anomaly segmentation is a critical task for driving applications, and it is approached traditionally as a per-pixel classification problem. However, reasoning individually about each pixel without considering their contextual semantics results in high uncertainty around the objects’ boundaries and numerous false positives. We propose a paradigm change by shifting from a per-pixel classification to a mask classification. Our mask-based method, Mask2Anomaly, demonstrates the feasibility of integrating an anomaly detection method in a mask-classification architecture. Mask2Anomaly includes several technical novelties that are designed to improve the detection of anomalies in masks: i) a global masked attention module to focus individually on the foreground and background regions; ii) a mask contrastive learning that maximizes the margin between an anomaly and known classes; and iii) a mask refinement solution to reduce false positives. Mask2Anomaly achieves new state-of-the-art results across a range of benchmarks, both in the per-pixel and component-level evaluations. In particular, Mask2Anomaly reduces the average false positives rate by 60% w.r.t. the previous state-of-the-art.

Visual Place recognition is commonly addressed as an image retrieval problem. However, retrieval methods are impractical to scale to large datasets, densely sampled from city-wide maps, since their dimension impact negatively on the inference time. Using approximate nearest neighbour search for retrieval helps to mitigate this issue, at the cost of a performance drop. In this paper we investigate whether we can effectively approach this task as a classification problem, thus bypassing the need for a similarity search. We find that existing classification methods for coarse, planet-wide localization are not suitable for the fine-grained and city-wide setting. This is largely due to how the dataset is split into classes, because these methods are designed to handle a sparse distribution of photos and as such do not consider the visual aliasing problem across neighbouring classes that naturally arises in dense scenarios. Thus, we propose a partitioning scheme that enables a fast and accurate inference, preserving a simple learning procedure, and a novel inference pipeline based on an ensemble of novel classifiers that uses the prototypes learned via an angular margin loss. Our method, Divide&Classify (D&C), enjoys the fast inference of classification solutions and an accuracy competitive with retrieval methods on the fine-grained, city-wide setting. Moreover, we show that D&C can be paired with existing retrieval pipelines to speed up computations by over 20 times while increasing their recall, leading to new state-of-the-art results. Code is available at https://github.com/ga1i13o/Divide-and-Classify

In the last decade, most research in Machine Learning contributed to the improvement of existing models, with the aim of increasing the performance of neural networks for the solution of a variety of different tasks. However, such advancements often come at the cost of an increase of model memory and computational requirements. This represents a significant limitation for the deployability of research output in realistic settings, where the cost, the energy consumption, and the complexity of the framework play a crucial role. To solve this issue, the designer should search for models that maximise the performance while limiting its footprint. Typical approaches to reach this goal rely either on manual procedures, which cannot guarantee the optimality of the final design, or upon Neural Architecture Search algorithms to automatise the process, at the expenses of extremely high computational time. This paper provides a solution for the fast identification of a neural network that maximises the model accuracy while preserving size and computational constraints typical of tiny devices. Our approach, named FreeREA, is a custom cell-based evolution NAS algorithm that exploits an optimised combination of training-free metrics to rank architectures during the search, thus without need of model training. Our experiments, carried out on the common benchmarks NAS-Bench-101 and NATS-Bench, demonstrate that i) FreeREA is a fast, efficient, and effective search method for models automatic design; ii) it outperforms State of the Art training-based and training-free techniques in all the datasets and benchmarks considered, and iii) it can easily generalise to constrained scenarios, representing a competitive solution for fast Neural Architecture Search in generic constrained applications. The code is available at https://github.com/NiccoloCavagnero/FreeREA.

Federated Learning (FL) has recently emerged as a possible way to tackle the domain shift in real-world Semantic Segmentation (SS) without compromising the private nature of the collected data. However, most of the existing works on FL unrealistically assume labeled data in the re-mote clients. Here we propose a novel task (FFreeDA) in which the clients’ data is unlabeled and the server accesses a source labeled dataset for pre-training only. To solve FFreeDA, we propose LADD, which leverages the knowledge of the pre-trained model by employing self-supervision with ad-hoc regularization techniques for local training and introducing a novel federated clustered aggregation scheme based on the clients’ style. Our experiments show that our algorithm is able to efficiently tackle the new task out-performing existing approaches. The code is available at https://github.com/Erosinho13/LADD.

Visual Geo-localization (VG) is the task of estimating the position where a given photo was taken by comparing it with a large database of images of known locations. To investigate how existing techniques would perform on a real-world city-wide VG application, we build San Francisco eXtra Large, a new dataset covering a whole city and providing a wide range of challenging cases, with a size 30x bigger than the previous largest dataset for visual geo-localization. We find that current methods fail to scale to such large datasets, therefore we design a new highly scalable training technique, called CosPlace, which casts the training as a classification problem avoiding the expensive mining needed by the commonly used contrastive learning. We achieve state-of-the-art performance on a wide range of datasets and find that CosPlace is robust to heavy domain changes. Moreover, we show that, compared to the previous state-of-the-art, CosPlace requires roughly 80% less GPU memory at train time, and it achieves better results with 8x smaller descriptors, paving the way for city-wide real-world visual geo-localization. Dataset, code and trained models are available for research purposes at https://github.com/gmberton/CosPlace.

In this paper, we propose a new open-source benchmarkingframeworkfor Visual Geo-localization (VG) that allows to build, train, and test a wide range of commonly used ar-chitectures, with the flexibility to change individual components of a geo-localization pipeline. The purpose of this framework is twofold: i) gaining insights into how differ-ent components and design choices in a VG pipeline im-pact the final results, both in terms of performance (re-call@N metric) and system requirements (such as execution time and memory consumption); ii) establish a system-atic evaluation protocol for comparing different methods. Using the proposed framework, we perform a large suite of experiments which provide criteria for choosing back-bone, aggregation and negative mining depending on the use-case and requirements. We also assess the impact of engineering techniques like pre/post-processing, data aug-mentation and image resizing, showing that better performance can be obtained through somewhat simple procedures: for example, downscaling the images' resolution to 80% can lead to similar results with a 36% savings in ex-traction time and dataset storage requirement. Code and trained models are available at dataset storage requirement. https://deep-vg-bench.herokuapp.com/.

Although existing semantic segmentation approaches achieve impressive results, they still struggle to update their models incrementally as new categories are uncovered. Furthermore, pixel-by-pixel annotations are expensive and time-consuming. This paper proposes a novel framework for Weakly Incremental Learning for Semantic Segmentation, that aims at learning to segment new classes from cheap and largely available image-level labels. As opposed to existing approaches, that need to generate pseudolabels offline, we use a localizer, trained with image-level labels and regularized by the segmentation model, to obtain pseudo-supervision online and update the model incrementally. We cope with the inherent noise in the process by using soft-labels generated by the localizer. We demonstrate the effectiveness of our approach on the Pascal VOC and COCO datasets, outperforming offline weakly-supervised methods and obtaining results comparable with incremental learning methods with full supervision.11Code can be found at https://github.com/fcd194/WILSON.

Event cameras are novel bio-inspired sensors, which asynchronously capture pixel-level intensity changes in the form of “events”. Due to their sensing mechanism, event cameras have little to no motion blur, a very high temporal resolution and require significantly less power and memory than traditional frame-based cameras. These characteristics make them a perfect fit to several real-world applications such as egocentric action recognition on wearable devices, where fast camera motion and limited power challenge traditional vision sensors. However, the ever-growing field of event-based vision has, to date, overlooked the potential of event cameras in such applications. In this paper, we show that event data is a very valuable modality for egocentric action recognition. To do so, we introduce N-EPIC-Kitchens, the first event-based camera extension of the large-scale EPIC-Kitchens dataset. In this context, we propose two strategies: (i) directly processing event-camera data with traditional video-processing architectures (E2(GO)) and (ii) using event-data to distill optical flow information (E2(GO)MO). On our proposed benchmark, we show that event data provides a comparable performance to RGB and optical flow, yet without any additional flow computation at deploy time, and an improved performance of up to 4% with respect to RGB only information. The N-EPIC-Kitchens dataset is available at https://github.com/EgocentricVision/N-EPIC-Kitchens.

Federated Learning (FL) allows training machine learning models in privacy-constrained scenarios by enabling the cooperation of edge devices without requiring local data sharing. This approach raises several challenges due to the different statistical distribution of the local datasets and the clients’ computational heterogeneity. In particular, the presence of highly non-i.i.d. data severely impairs both the performance of the trained neural network and its convergence rate, increasing the number of communication rounds requested to reach a performance comparable to that of the centralized scenario. As a solution, we propose FedSeq, a novel framework leveraging the sequential training of subgroups of heterogeneous clients, i.e. superclients, to emulate the centralized paradigm in a privacy-compliant way. Given a fixed budget of communication rounds, we show that FedSeq outperforms or match several state-of-the-art federated algorithms in terms of final performance and speed of convergence. Finally, our method can be easily integrated with other approaches available in the literature. Empirical results show that combining existing algorithms with FedSeq further improves its final performance and convergence speed. We test our method on CIFAR-10 and CIFAR-100 and prove its effectiveness in both i.i.d. and non-i.i.d. scenarios.1

Semantic Segmentation is essential to make self-driving vehicles autonomous, enabling them to understand their surroundings by assigning individual pixels to known categories. However, it operates on sensible data collected from the users' cars; thus, protecting the clients' privacy becomes a primary concern. For similar reasons, Federated Learning has been recently introduced as a new machine learning paradigm aiming to learn a global model while preserving privacy and leveraging data on millions of remote devices. Despite several efforts on this topic, no work has explicitly addressed the challenges of federated learning in semantic segmentation for driving so far. To fill this gap, we propose FedDrive, a new benchmark consisting of three settings and two datasets, incorporating the real-world challenges of statistical heterogeneity and domain generalization. We benchmark state-of-the-art algorithms from the federated learning literature through an in-depth analysis, combining them with style transfer methods to improve their generalization ability. We demonstrate that correctly handling normalization statistics is crucial to deal with the aforementioned challenges. Furthermore, style transfer improves performance when dealing with significant appearance shifts. Official website: https://feddrive.github.io.

Vision systems trained in closed-world scenarios fail when presented with new environmental conditions, new data distributions, and novel classes at deployment time. How to move towards open-world learning is a long-standing research question. The existing solutions mainly focus on specific aspects of the problem (single domain Open-Set, multi-domain Closed-Set), or propose complex strategies which combine several losses and manually tuned hyperparameters. In this work, we tackle multi-source Open-Set domain adaptation by introducing HyMOS: a straightforward model that exploits the power of contrastive learning and the properties of its hyperspherical feature space to correctly predict known labels on the target, while rejecting samples belonging to any unknown class. Hy-MOS includes style transfer among the instance transformations of contrastive learning to get domain invariance while avoiding the risk of negative-transfer. A self-paced threshold is defined on the basis of the observed data distribution and updates online during training, allowing to handle the known-unknown separation. We validate our method over three challenging datasets. The obtained results show that HyMOS outperforms several competitors, defining the new state-of-the-art. Our code is available at https://github.com/silvia1993/HyMOS.

First person action recognition is becoming an increasingly researched area thanks to the rising popularity of wearable cameras. This is bringing to light cross-domain issues that are yet to be addressed in this context. Indeed, the information extracted from learned representations suffers from an intrinsic "environmental bias". This strongly affects the ability to generalize to unseen scenarios, limiting the application of current methods to real settings where labeled data are not available during training. In this work, we introduce the first domain generalization approach for egocentric activity recognition, by proposing a new audiovisual loss, called Relative Norm Alignment loss. It rebalances the contributions from the two modalities during training, over different domains, by aligning their feature norm representations. Our approach leads to strong results in domain generalization on both EPIC-Kitchens-55 and EPIC-Kitchens-100, as demonstrated by extensive experiments, and can be extended to work also on domain adaptation settings with competitive results.

Unsupervised Domain Adaptation (UDA) is a key issue in visual recognition, as it allows to bridge different visual domains enabling robust performances in the real world. To date, all proposed approaches rely on human expertise to manually adapt a given UDA method (e.g. DANN) to a specific backbone architecture (e.g. ResNet). This dependency on handcrafted designs limits the applicability of a given approach in time, as old methods need to be constantly adapted to novel backbones.Existing Neural Architecture Search (NAS) approaches cannot be directly applied to mitigate this issue, as they rely on labels that are not available in the UDA setting. Furthermore, most NAS methods search for full architectures, which precludes the use of pre-trained models, essential in a vast range of UDA settings for reaching SOTA results.To the best of our knowledge, no prior work has addressed these aspects in the context of NAS for UDA. Here we tackle both aspects with an Adversarial Branch Architecture Search for UDA (ABAS): i. we address the lack of target labels by a novel data-driven ensemble approach for model selection; and ii. we search for an auxiliary adversarial branch, attached to a pre-trained backbone, which drives the domain alignment. We extensively validate ABAS to improve two modern UDA techniques, DANN and ALDA, on three standard visual recognition datasets (Office31, Office-Home and PACS). In all cases, ABAS robustly finds the adversarial branch architectures and parameters which yield best performances. https://github.com/lr94/abas.

In this paper we consider the task of semantic segmentation in autonomous driving applications. Specifically, we consider the cross-domain few-shot setting where training can use only few real-world annotated images and many annotated synthetic images. In this context, aligning the domains is made more challenging by the pixel-wise class imbalance that is intrinsic in the segmentation and that leads to ignoring the underrepresented classes and overfitting the well represented ones. We address this problem with a novel framework called Pixel-By-Pixel Cross-Domain Alignment (PixDA). We propose a novel pixel-by-pixel domain adversarial loss following three criteria: (i) align the source and the target domain for each pixel, (ii) avoid negative transfer on the correctly represented pixels, and (iii) regularize the training of infrequent classes to avoid overfitting. The pixel-wise adversarial training is assisted by a novel sample selection procedure, that handles the imbalance between source and target data, and a knowledge distillation strategy, that avoids overfitting towards the few target images. We demonstrate on standard synthetic-to-real benchmarks that PixDA outperforms previous state-of-the-art methods in (1-5)-shot settings.1

Human adaptability relies crucially on learning and merging knowledge from both supervised and unsupervised tasks: the parents point out few important concepts, but then the children fill in the gaps on their own. This is particularly effective, because supervised learning can never be exhaustive and thus learning autonomously allows to discover invariances and regularities that help to generalize. In this paper we propose to apply a similar approach to the problem of object recognition across domains: our model learns the semantic labels in a supervised fashion, and broadens its understanding of the data by learning from self-supervised signals on the same images. This secondary task helps the network to focus on object shapes, learning concepts like spatial orientation and part correlation, while acting as a regularizer for the classification task over multiple visual domains. Extensive experiments confirm our intuition and show that our multi-task method, combining supervised and self-supervised knowledge, provides competitive results with respect to more complex domain generalization and adaptation solutions. It also proves its potential in the novel and challenging predictive and partial domain adaptation scenarios.

Deep neural networks have enabled major progresses in semantic segmentation. However, even the most advanced neural architectures suffer from important limitations. First, they are vulnerable to catastrophic forgetting, i.e., they perform poorly when they are required to incrementally update their model as new classes are available. Second, they rely on large amount of pixel-level annotations to produce accurate segmentation maps. To tackle these issues, we introduce a novel incremental class learning approach for semantic segmentation taking into account a peculiar aspect of this task: since each training step provides annotation only for a subset of all possible classes, pixels of the background class exhibit a semantic shift. Therefore, we revisit the traditional distillation paradigm by designing novel loss terms which explicitly account for the background shift. Additionally, we introduce a novel strategy to initialize classifier's parameters at each step in order to prevent biased predictions toward the background class. Finally, we demonstrate that our approach can be extended to point- and scribble-based weakly supervised segmentation, modeling the partial annotations to create priors for unlabeled pixels. We demonstrate the effectiveness of our approach with an extensive evaluation on the Pascal-VOC, ADE20K, and Cityscapes datasets, significantly outperforming state-of-the-art methods.

In this paper we propose a novel method for image matching based on dense local features and tailored for visual geolocalization. Dense local features matching is robust against changes in illumination and occlusions, but not against viewpoint shifts which are a fundamental aspect of geolocalization. Our method, called GeoWarp, directly embeds invariance to viewpoint shifts in the process of extracting dense features. This is achieved via a trainable module which learns from the data an invariance that is meaningful for the task of recognizing places. We also devise a new self-supervised loss and two new weakly supervised losses to train this module using only unlabeled data and weak labels. GeoWarp is implemented efficiently as a re-ranking method that can be easily embedded into pre-existing visual geolocalization pipelines. Experimental validation on standard geolocalization benchmarks demonstrates that GeoWarp boosts the accuracy of state-of-the-art retrieval architectures. The code and trained models will be released upon acceptance of this paper.

We address the task of cross-domain visual place recognition, where the goal is to geolocalize a given query image against a labeled gallery, in the case where the query and the gallery belong to different visual domains. To achieve this, we focus on building a domain robust deep network by leveraging over an attention mechanism combined with few-shot unsupervised domain adaptation techniques, where we use a small number of unlabeled target domain images to learn about the target distribution. With our method, we are able to outperform the current state of the art while using two orders of magnitude less target domain images. Finally we propose a new large-scale dataset for cross-domain visual place recognition, called SVOX. The pytorch code is available at https://github.com/valeriopaolicelli/AdAGeo.

One of the main challenges for developing visual recognition systems working in the wild is to devise computational models immune from the domain shift problem, i.e., accurate when test data are drawn from a (slightly) different data distribution than training samples. In the last decade, several research efforts have been devoted to devise algorithmic solutions for this issue. Recent attempts to mitigate domain shift have resulted into deep learning models for domain adaptation which learn domain-invariant representations by introducing appropriate loss terms, by casting the problem within an adversarial learning framework or by embedding into deep network specific domain normalization layers. This paper describes a novel approach for unsupervised domain adaptation. Similarly to previous works we propose to align the learned representations by embedding them into appropriate network feature normalization layers. Opposite to previous works, our Domain Alignment Layers are designed not only to match the source and target feature distributions but also to automatically learn the degree of feature alignment required at different levels of the deep network. Differently from most previous deep domain adaptation methods, our approach is able to operate in a multi-source setting. Thorough experiments on four publicly available benchmarks confirm the effectiveness of our approach.

Unsupervised Domain Adaptation (UDA) refers to the problem of learning a model in a target domain where labeled data are not available by leveraging information from annotated data in a source domain. Most deep UDA approaches operate in a single-source, single-target scenario, i.e., they assume that the source and the target samples arise from a single distribution. However, in practice most datasets can be regarded as mixtures of multiple domains. In these cases, exploiting traditional single-source, single-target methods for learning classification models may lead to poor results. Furthermore, it is often difficult to provide the domain labels for all data points, i.e. latent domains should be automatically discovered. This paper introduces a novel deep architecture which addresses the problem of UDA by automatically discovering latent domains in visual datasets and exploiting this information to learn robust target classifiers. Specifically, our architecture is based on two main components, i.e. a side branch that automatically computes the assignment of each sample to its latent domain and novel layers that exploit domain membership information to appropriately align the distribution of the CNN internal feature representations to a reference distribution. We evaluate our approach on publicly available benchmarks, showing that it outperforms state-of-the-art domain adaptation methods.

Despite their effectiveness in a wide range of tasks, deep architectures suffer from some important limitations. In particular, they are vulnerable to catastrophic forgetting, i.e. they perform poorly when they are required to update their model as new classes are available but the original training set is not retained. This paper addresses this problem in the context of semantic segmentation. Current strategies fail on this task because they do not consider a peculiar aspect of semantic segmentation: since each training step provides annotation only for a subset of all possible classes, pixels of the background class (i.e. pixels that do not belong to any other classes) exhibit a semantic distribution shift. In this work we revisit classical incremental learning methods, proposing a new distillation-based framework which explicitly accounts for this shift. Furthermore, we introduce a novel strategy to initialize classifier's parameters, thus preventing biased predictions toward the background class. We demonstrate the effectiveness of our approach with an extensive evaluation on the Pascal-VOC 2012 and ADE20K datasets, significantly outperforming state of the art incremental learning methods.

Wearable cameras are becoming more and more popular in several applications, increasing the interest of the research community in developing approaches for recognizing actions from the first-person point of view. An open challenge in egocentric action recognition is that videos lack detailed information about the main actor's pose and thus tend to record only parts of the movement when focusing on manipulation tasks. Thus, the amount of information about the action itself is limited, making crucial the understanding of the manipulated objects and their context. Many previous works addressed this issue with two-stream architectures, where one stream is dedicated to modeling the appearance of objects involved in the action, and another to extracting motion features from optical flow. In this paper, we argue that learning features jointly from these two information channels is beneficial to capture the spatio-temporal correlations between the two better. To this end, we propose a single stream architecture able to do so, thanks to the addition of a self-supervised block that uses a pretext motion prediction task to intertwine motion and appearance knowledge. Experiments on several publicly available databases show the power of our approach.

Recent unsupervised domain adaptation methods based on deep architectures have shown remarkable performance not only in traditional classification tasks but also in more complex problems involving structured predictions (e.g. semantic segmentation, depth estimation). Following this trend, in this paper we present a novel deep adaptation framework for estimating keypoints under domain shift, i.e, when the training (source) and the test (target) images significantly differ in terms of visual appearance. Our method seamlessly combines three different components: feature alignment, adversarial training and self-supervision. Specifically, our deep architecture leverages from domain-specific distribution alignment layers to perform target adaptation at the feature level. Furthermore, a novel loss is proposed which combines an adversarial term for ensuring aligned predictions in the output space and a geometric consistency term which guarantees coherent predictions between a target sample and its perturbed version. Our extensive experimental evaluation conducted on three publicly available benchmarks shows that our approach outperforms state-of-the-art domain adaptation methods in the 2D keypoint prediction task.

Motivated by recent advances in deep domain adaptation, this paper introduces a deep architecture for estimating 3D keypoints when the training (source) and the test (target) images greatly differ in terms of visual appearance (domain shift). Our approach operates by promoting domain distribution alignment in the feature space adopting batch normalization-based techniques. Furthermore, we propose to collect statistics about 3D keypoints positions of the source training data and to use this prior information to constrain predictions on the target domain introducing a loss derived from Multidimensional Scaling. We conduct an extensive experimental evaluation considering three publicly available benchmarks and show that our approach out-performs state-of-the-art domain adaptation methods for 3D keypoints predictions.

Human adaptability relies crucially on the ability to learn and merge knowledge both from supervised and unsupervised learning: the parents point out few important concepts, but then the children fill in the gaps on their own. This is particularly effective, because supervised learning can never be exhaustive and thus learning autonomously allows to discover invariances and regularities that help to generalize. In this paper we propose to apply a similar approach to the task of object recognition across domains: our model learns the semantic labels in a supervised fashion, and broadens its understanding of the data by learning from self-supervised signals how to solve a jigsaw puzzle on the same images. This secondary task helps the network to learn the concepts of spatial correlation while acting as a regularizer for the classification task. Multiple experiments on the PACS, VLCS, Office-Home and digits datasets confirm our intuition and show that this simple method outperforms previous domain generalization and adaptation solutions. An ablation study further illustrates the inner workings of our approach.

The ability to categorize is a cornerstone of visual intelligence, and a key functionality for artificial, autonomous visual machines. This problem will never be solved without algorithms able to adapt and generalize across visual domains. Within the context of domain adaptation and generalization, this paper focuses on the predictive domain adaptation scenario, namely the case where no target data are available and the system has to learn to generalize from annotated source images plus unlabeled samples with associated metadata from auxiliary domains. Our contribution is the first deep architecture that tackles predictive domain adaptation, able to leverage over the information brought by the auxiliary domains through a graph. Moreover, we present a simple yet effective strategy that allows us to take advantage of the incoming target data at test time, in a continuous domain adaptation scenario. Experiments on three benchmark databases support the value of our approach.

While today's robots are able to perform sophisticated tasks, they can only act on objects they have been trained to recognize. This is a severe limitation: any robot will inevitably see new objects in unconstrained settings, and thus will always have visual knowledge gaps. However, standard visual modules are usually built on a limited set of classes and are based on the strong prior that an object must belong to one of those classes. Identifying whether an instance does not belong to the set of known categories (i.e. open set recognition), only partially tackles this problem, as a truly autonomous agent should be able not only to detect what it does not know, but also to extend dynamically its knowledge about the world. We contribute to this challenge with a deep learning architecture that can dynamically update its known classes in an end-to-end fashion. The proposed deep network, based on a deep extension of a non-parametric model, detects whether a perceived object belongs to the set of categories known by the system and learns it without the need to retrain the whole system from scratch. Annotated images about the new category can be provided by an `oracle' (i.e. human supervision), or by autonomous mining of the Web. Experiments on two different databases and on a robot platform demonstrate the promise of our approach.

Deep networks have brought significant advances in robot perception, enabling to improve the capabilities of robots in several visual tasks, ranging from object detection and recognition to pose estimation, semantic scene segmentation and many others. Still, most approaches typically address visual tasks in isolation, resulting in overspecialized models which achieve strong performances in specific applications but work poorly in other (often related) tasks. This is clearly sub-optimal for a robot which is often required to perform simultaneously multiple visual recognition tasks in order to properly act and interact with the environment. This problem is exacerbated by the limited computational and memory resources typically available onboard to a robotic platform. The problem of learning flexible models which can handle multiple tasks in a lightweight manner has recently gained attention in the computer vision community and benchmarks supporting this research have been proposed. In this work we study this problem in the robot vision context, proposing a new benchmark, the RGB-D Triathlon, and evaluating state of the art algorithms in this novel challenging scenario. We also define a new evaluation protocol, better suited to the robot vision setting. Results shed light on the strengths and weaknesses of existing approaches and on open issues, suggesting directions for future research.

Current Domain Adaptation (DA) methods based on deep architectures assume that the source samples arise from a single distribution. However, in practice most datasets can be regarded as mixtures of multiple domains. In these cases exploiting single-source DA methods for learning target classifiers may lead to sub-optimal, if not poor, results. In addition, in many applications it is difficult to manually provide the domain labels for all source data points, i.e. latent domains should be automatically discovered. This paper introduces a novel Convolutional Neural Network (CNN) architecture which (i) automatically discovers latent domains in visual datasets and (ii) exploits this information to learn robust target classifiers. Our approach is based on the introduction of two main components, which can be embedded into any existing CNN architecture: (i) a side branch that automatically computes the assignment of a source sample to a latent domain and (ii) novel layers that exploit domain membership information to appropriately align the distribution of the CNN internal feature representations to a reference distribution. We test our approach on publicly-available datasets, showing that it outperforms state-of-the-art multi-source DA methods by a large margin.

The effectiveness of GANs in producing images according to a specific visual domain has shown potential in unsupervised domain adaptation. Source labeled images have been modified to mimic target samples for training classifiers in the target domain, and inverse mappings from the target to the source domain have also been evaluated, without new image generation. In this paper we aim at getting the best of both worlds by introducing a symmetric mapping among domains. We jointly optimize bi-directional image transformations combining them with target self-labeling. We define a new class consistency loss that aligns the generators in the two directions, imposing to preserve the class identity of an image passing through both domain mappings. A detailed analysis of the reconstructed images, a thorough ablation study and extensive experiments on six different settings confirm the power of our approach.

A long standing problem in visual object categorization is the ability of algorithms to generalize across different testing conditions. The problem has been formalized as a covariate shift among the probability distributions generating the training data (source) and the test data (target) and several domain adaptation methods have been proposed to address this issue. While these approaches have considered the single source-single target scenario, it is plausible to have multiple sources and require adaptation to any possible target domain. This last scenario, named Domain Generalization (DG), is the focus of our work. Differently from previous DG methods which learn domain invariant representations from source data, we design a deep network with multiple domain-specific classifiers, each associated to a source domain. At test time we estimate the probabilities that a target sample belongs to each source domain and exploit them to optimally fuse the classifiers predictions. To further improve the generalization ability of our model, we also introduced a domain agnostic component supporting the final classifier. Experiments on two public benchmarks demonstrate the power of our approach.

A commercial robot, trained by its manufacturer to recognize a predefined number and type of objects, might be used in many settings, that will in general differ in their illumination conditions, background, type and degree of clutter, and so on. Recent computer vision works tackle this generalization issue through domain adaptation methods, assuming as source the visual domain where the system is trained and as target the domain of deployment. All approaches assume to have access to images from all classes of the target during training, an unrealistic condition in robotics applications. We address this issue proposing an algorithm that takes into account the specific needs of robot vision. Our intuition is that the nature of the domain shift experienced mostly in robotics is local. We exploit this through the learning of maps that spatially ground the domain and quantify the degree of shift, embedded into an end-to-end deep domain adaptation architecture. By explicitly localizing the roots of the domain shift we significantly reduce the number of parameters of the architecture to tune, we gain the flexibility necessary to deal with subset of categories in the target domain at training time, and we provide a clear feedback on the rationale behind any classification decision, which can be exploited in human-robot interactions. Experiments on two different settings of the iCub World database confirm the suitability of our method for robot vision.

The ability to recognize objects is an essential skill for a robotic system acting in human-populated environments. Despite decades of effort from the robotic and vision research communities, robots are still missing good visual perceptual systems, preventing the use of autonomous agents for realworld applications. The progress is slowed down by the lack of a testbed able to accurately represent the world perceived by the robot in-the-wild. In order to fill this gap, we introduce a large-scale, multi-view object dataset collected with an RGB-D camera mounted on a mobile robot. The dataset embeds the challenges faced by a robot in a real-life application and provides a useful tool for validating object recognition algorithms. Besides describing the characteristics of the dataset, the paper evaluates the performance of a collection of well-established deep convolutional networks on the new dataset and analyzes the transferability of deep representations from Web images to robotic data. Despite the promising results obtained with such representations, the experiments demonstrate that object classification with real-life robotic data is far from being solved. Finally, we provide a comparative study to analyze and highlight the open challenges in robot vision, explaining the discrepancies in the performance.

Technological developments call for increasing perception and action capabilities of robots. Among other skills, vision systems that can adapt to any possible change in the working conditions are needed. Since these conditions are unpredictable, we need benchmarks which allow to assess the generalization and robustness capabilities of our visual recognition algorithms. In this work we focus on robotic kitting in unconstrained scenarios. As a first contribution, we present a new visual dataset for the kitting task. Differently from standard object recognition datasets, we provide images of the same objects acquired under various conditions where camera, illumination and background are changed. This novel dataset allows for testing the robustness of robot visual recognition algorithms to a series of different domain shifts both in isolation and unified. Our second contribution is a novel online adaptation algorithm for deep models, based on batch-normalization layers, which allows to continuously adapt a model to the current working conditions. Differently from standard domain adaptation algorithms, it does not require any image from the target domain at training time. We benchmark the performance of the algorithm on the proposed dataset, showing its capability to fill the gap between the performances of a standard architecture and its counterpart adapted offline to the given target domain.

Classifiers trained on given databases perform poorly when tested on data acquired in different settings. This is explained in domain adaptation through a shift among distributions of the source and target domains. Attempts to align them have traditionally resulted in works reducing the domain shift by introducing appropriate loss terms, measuring the discrepancies between source and target distributions, in the objective function. Here we take a different route, proposing to align the learned representations by embedding in any given network specific Domain Alignment Layers, designed to match the source and target feature distributions to a reference one. Opposite to previous works which define a priori in which layers adaptation should be performed, our method is able to automatically learn the degree of feature alignment required at different levels of the deep network. Thorough experiments on different public benchmarks, in the unsupervised setting, confirm the power of our approach.

Convolutional Neural Networks (CNNs) trained on large scale RGB databases have become the secret sauce in the majority of recent approaches for object categorization from RGB-D data. Thanks to colorization techniques, these methods exploit the filters learned from 2D images to extract meaningful representations in 2.5D. Still, the perceptual signature of these two kind of images is very different, with the first usually strongly characterized by textures, and the second mostly by silhouettes of objects. Ideally, one would like to have two CNNs, one for RGB and one for depth, each trained on a suitable data collection, able to capture the perceptual properties of each channel for the task at hand. This has not been possible so far, due to the lack of a suitable depth database. This paper addresses this issue, proposing to opt for synthetically generated images rather than collecting by hand a 2.5D large scale database. While being clearly a proxy for real data, synthetic images allow to trade quality for quantity, making it possible to generate a virtually infinite amount of data. We show that the filters learned from such data collection, using the very same architecture typically used on visual data, learns very different filters, resulting in depth features (a) able to better characterize the different facets of depth images, and (b) complementary with respect to those derived from CNNs pre-trained on 2D datasets. Experiments on two publicly available databases show the power of our approach.

Autonomous robots that are to assist humans in their daily lives must recognize and understand the meaning of objects in their environment. However, the open nature of the world means robots must be able to learn and extend their knowledge about previously unknown objects on-line. In this work we investigate the problem of unknown object hypotheses generation, and employ a semantic Web-mining framework along with deep-learning-based object detectors. This allows us to make use of both visual and semantic features in combined hypotheses generation. Experiments on data from mobile robots in real world application deployments show that this combination improves performance over the use of either method in isolation.

Deep networks thrive when trained on large scale data collections. This has given ImageNet a central role in the development of deep architectures for visual object classification. However, ImageNet was created during a specific period in time, and as such it is prone to aging, as well as dataset bias issues. Moving beyond fixed training datasets will lead to more robust visual systems, especially when deployed on robots in new environments which must train on the objects they encounter there. To make this possible, it is important to break free from the need for manual annotators. Recent work has begun to investigate how to use the massive amount of images available on the Web in place of manual image annotations. We contribute to this research thread with two findings: (1) a study correlating a given level of noisily labels to the expected drop in accuracy, for two deep architectures, on two different types of noise, that clearly identifies GoogLeNet as a suitable architecture for learning from Web data; (2) a recipe for the creation of Web datasets with minimal noise and maximum visual variability, based on a visual and natural language processing concept expansion strategy. By combining these two results, we obtain a method for learning powerful deep object models automatically from the Web. We confirm the effectiveness of our approach through object categorization experiments using our Web-derived version of ImageNet on a popular robot vision benchmark database, and on a lifelong object discovery task on a mobile robot.

Since Convolutional Neural Networks (CNNs) have become the leading learning paradigm in visual recognition, Naive Bayes Nearest Neighbor (NBNN)-based classifiers have lost momentum in the community. This is because (1) such algorithms cannot use CNN activations as input features, (2) they cannot be used as final layer of CNN architectures for end-to-end training, and (3) they are generally not scalable and hence cannot handle big data. This paper proposes a framework that addresses all these issues, thus bringing back NBNNs on the map. We solve the first by extracting CNN activations from local patches at multiple scale levels, similarly to [13]. We address simultaneously the second and third by proposing a scalable version of Naive Bayes Non-linear Learning (NBNL, [7]). Results obtained using pre-trained CNNs on standard scene and domain adaptation databases show the strength of our approach, opening a new season for NBNNs.

The transfer learning and domain adaptation problems originate from a distribution mismatch between the source and target data distribution. The causes of such mismatch are traditionally considered different. Thus, transfer learn- ing and domain adaptation algorithms are designed to ad- dress different issues, and cannot be used in both settings unless substantially modified. Still, one might argue that these problems are just different declinations of learning to learn, i.e. the ability to leverage over prior knowledge when attempting to solve a new task. We propose a learning to learn framework able to lever- age over source data regardless of the origin of the distri- bution mismatch. We consider prior models as experts, and use their output confidence value as features. We use them to build the new target model, combined with the features from the target data through a high-level cue integration scheme. This results in a class of algorithms usable in a plug-and-play fashion over any learning to learn scenario, from binary and multi-class transfer learning to single and multiple source domain adaptation settings. Experiments on several public datasets show that our approach consis- tently achieves the state of the art.

A crucial feature of a good scene recognition algorithm is its ability to generalize. Scene categories, especially those related to human made indoor places or to human activities like sports, do present a high degree of intra-class variability, which in turn requires high robustness and generalization properties. Such features are amongst the distinctive characteristics of the Naive Bayes Nearest Neighbor (NBNN) approach [1], an image classification framework that since its introduction in 2008 has been gaining momentum in the visual recognition community. In this paper we show how with a straightforward modification of the original NBNN scoring function it is possible to use a recently introduced latent locally linear SVM algorithm to discriminatively learn a set of prototype local features for each class. The resulting classification algorithm, that we call Naive Bayes Nonlinear Learning (NBNL) preserves the generality and robustness properties of the original approach, while greatly reducing its memory requirements during testing, and significantly improving its performance. To the best of our knowledge this is the first work to exploit the structure of the local features through the use of a latent locally linear discriminative learning method. Experiments over three different public scene recognition datasets show the effectiveness of the proposed algorithm, which outperforms several existing NBNN-based methods and is competitive with standard Bag-of-Words plus SVM approaches.

We present a benchmark of several existing multi-source adaptive methods on the largest publicly available database of surface electromyography signals for polyarticulated self-powered hand prostheses. By exploiting the information collected over numerous subjects, these methods allow to reduce significantly the training time needed by any new prosthesis user. Our findings provide the bio robotics community with a deeper understanding of adaptive learning solutions for user-machine control and pave the way for further improvements in hand-prosthetics.

Learning a visual object category from few samples is a compelling and challenging problem. In several real-world applications collecting many annotated data is costly and not always possible. However, a small training set does not allow to cover the high intraclass variability typical of visual objects. In this condition, machine learning methods provide very few guarantees. This paper presents a discriminative model adaptation algorithm able to proficiently learn a target object with few examples by relying on other previously learned source categories. The proposed method autonomously chooses from where and how much to transfer information by solving a convex optimization problem which ensures to have the minimal leave-one-out error on the available training set. We analyze several properties of the described approach and perform an extensive experimental comparison with other existing transfer solutions, consistently showing the value of our algorithm.

Kernelized Support Vector Machines (SVM) have gained the status of off-the-shelf classifiers, able to deliver state of the art performance on almost any problem. Still, their practical use is constrained by their computational and memory complexity, which grows super-linearly with the number of training samples. In order to retain the low training and testing complexity of linear classifiers and the flexibility of non linear ones, a growing, promising alternative is represented by methods that learn non-linear classifiers through local combinations of linear ones. In this paper we propose a new multi class local classifier, based on a latent SVM formulation. The proposed classifier makes use of a set of linear models that are linearly combined using sample and class specific weights. Thanks to the latent formulation, the combination coefficients are modeled as latent variables. We allow soft combinations and we provide a closed-form solution for their estimation, resulting in an efficient prediction rule. This novel formulation allows to learn in a principled way the sample specific weights and the linear classifiers, in a unique optimization problem, using a CCCP optimization procedure. Extensive experiments on ten standard UCI machine learning datasets, one large binary dataset, three character and digit recognition databases, and a visual place categorization dataset show the power of the proposed approach.

Since the seminal work of Thrun [16], the learning to learn paradigm has been defined as the ability of an agent to improve its performance at each task with experience, with the number of tasks. Within the object categorization domain, the visual learning community has actively declined this paradigm in the transfer learning setting. Almost all proposed methods focus on category detection problems, addressing how to learn a new target class from few samples by leveraging over the known source. But if one thinks of learning over multiple tasks, there is a need for multiclass transfer learning algorithms able to exploit previous source knowledge when learning a new class, while at the same time optimizing their overall performance. This is an open challenge for existing transfer learning algorithms. The contribution of this paper is a discriminative method that addresses this issue, based on a Least-Squares Support Vector Machine formulation. Our approach is designed to balance between transferring to the new class and preserving what has already been learned on the source models. Extensive experiments on subsets of publicly available datasets prove the effectiveness of our approach.

Over the last years, several authors have signaled that state of the art categorization methods fail to perform well when trained and tested on data from different databases. The general consensus in the literature is that this issue, known as domain adaptation and/or dataset bias, is due to a distribution mismatch between data collections. Methods addressing it go from max-margin classifiers to learning how to modify the features and obtain a more robust representation. The large majority of these works use BOW feature descriptors, and learning methods based on image-to-image distance functions. Following the seminal work of [6], in this paper we challenge these two assumptions. We experimentally show that using the NBNN classifier over existing domain adaptation databases achieves always very strong performances. We build on this result, and present an NBNN-based domain adaptation algorithm that learns iteratively a class metric while inducing, for each sample, a large margin separation among classes. To the best of our knowledge, this is the first work casting the domain adaptation problem within the NBNN framework. Experiments show that our method achieves the state of the art, both in the unsupervised and semi-supervised settings.

At the time of this writing, the main means of control for polyarticulated self-powered hand prostheses is surface electromyography (sEMG). In the clinical setting, data collected from two electrodes are used to guide the hand movements selecting among a finite number of postures. Machine learning has been applied in the past to the sEMG signal (not in the clinical setting) with interesting results, which provide more insight on how these data could be used to improve prosthetic functionality. Researchers have mainly concentrated so far on increasing the accuracy of sEMG classification and/or regression, but, in general, a finer control implies a longer training period. A desirable characteristic would be to shorten the time needed by a patient to learn how to use the prosthesis. To this aim, we propose here a general method to reuse past experience, in the form of models synthesized from previous subjects, to boost the adaptivity of the prosthesis. Extensive tests on databases recorded from healthy subjects in controlled and noncontrolled conditions reveal that the method significantly improves the results over the baseline nonadaptive case. This promising approach might be employed to pretrain a prosthesis before shipping it to a patient, leading to a shorter training phase.

Indoor scenes are characterized by a high intra-class variability, mainly due to the intrinsic variety of the objects in them, and to the drastic image variations due to (even small) view-point changes. One of the main trends in the literature has been to employ representations coupling statistical characterizations of the image, with a description of their spatial distribution. This is usually done by combining multiple representations of different image regions, most often using a fixed 44, or pyramidal image-partitioning scheme. While these encodings are able to capture the spatial regularities of the problem, they are unsuitable to handle its spatial variabilities. In this work we propose to complement a traditional spatial-encoding scheme with a bottom-up approach designed to discover visual-structures regardless of their exact position in the scene. To this end we use saliency maps to segment each image in two regions: The most and least salient 50%. This segmentation provides a description of images which is somehow related to the relative semantics of the discovered regions, complementing the canonical spatial description. We evaluated the proposed technique on three public scene recognition datasets. Our results prove this approach to be effective in the indoor scenario, while being also meaningful for other scene categorization tasks.

Open ended learning is a dynamic process based on the continuous analysis of new data, guided by past experience. On one side it is helpful to take advantage of prior knowledge when only few information on a new task is available (transfer learning). On the other, it is important to continuously update an existing model so to exploit the new incoming data, especially if their informative content is very different from what is already known (online learning). Until today these two aspects of the learning process have been tackled separately. In this paper we propose an algorithm that takes the best of both worlds: we consider a sequential learning setting, and we exploit the potentiality of knowledge transfer with a computationally cheap solution. At the same time, by relying on past experience we boost online learning to predict reliably on future problems. A theoretical analysis, coupled with extensive experiments, show that our approach performs well in terms of the online number of training mistakes, as well as in terms of performance on separate test sets.

We present a large scale database of images and captions, designed for supporting research on how to use captioned images from the Web for training visual classifiers. It consists of more than 125,000 images of celebrities from different fields downloaded from the Web. Each image is associated to its original text caption, extracted from the html page the image comes from. We coin it FAN-Large, for Face And Names Large scale database. Its size and deliberate high level of noise makes it to our knowledge the largest and most realistic database supporting this type of research. The dataset and its annotations are publicly available and can be obtained from http://www.vision.ee.ethz.ch/~calvin/fanlarge/. We report results on a thorough assessment of FAN-Large using several existing approaches for name-face association, and present and evaluate new contextual features derived from the caption. Our findings provide important cues on the strengths and limitations of existing approaches.

The vast majority of transfer learning methods proposed in the visual recognition domain over the last years addresses the problem of object category detection, assuming a strong control over the priors from which transfer is done. This is a strict condition, as it concretely limits the use of this type of approach in several settings: for instance, it does not allow in general to use off-the-shelf models as priors. Moreover, the lack of a multiclass formulation for most of the existing transfer learning algorithms prevents using them for object categorization problems, where their use might be beneficial, especially when the number of categories grows and it becomes harder to get enough annotated data for training standard learning methods. This paper presents a multiclass transfer learning algorithm that allows to take advantage of priors built over different features and with different learning methods than the one used for learning the new task. We use the priors as experts, and transfer their outputs to the new incoming samples as additional information. We cast the learning problem within the Multi Kernel Learning framework. The resulting formulation solves efficiently a joint optimization problem that determines from where and how much to transfer, with a principled multiclass formulation. Extensive experiments illustrate the value of this approach.

The ability of building robust semantic space representations of environments is crucial for the development of truly autonomous robots. This task, inherently connected with cognition, is traditionally achieved by training the robot with a supervised learning phase. We argue that the design of robust and autonomous systems would greatly benefit from adopting a semi-supervised online learning approach. Indeed, the support of open-ended, lifelong learning is fundamental in order to cope with the dazzling variability of the real world, and online learning provides precisely this kind of ability. Here we focus on the robot place recognition problem, and we present an online place classification algorithm that is able to detect gap in its own knowledge based on a confidence measure. For every incoming new image frame, the method is able to decide if (a) it is a known room with a familiar appearance, (b) it is a known room with a challenging appearance, or (c) it is a new, unknown room. Experiments on a subset of the challenging COLD database show the promise of our approach.

Several object categorization algorithms use kernel methods over multiple cues, as they offer a principled approach to combine multiple cues, and to obtain state-of-the-art performance. A general drawback of these strategies is the high computational cost during training, that prevents their application to large-scale problems. They also do not provide theoretical guarantees on their convergence rate. Here we present a Multiclass Multi Kernel Learning (MKL) algorithm that obtains state-of-the-art performance in a considerably lower training time. We generalize the standard MKL formulation to introduce a parameter that allows us to decide the level of sparsity of the solution. Thanks to this new setting, we can directly solve the problem in the primal formulation. We prove theoretically and experimentally that 1) our algorithm has a faster convergence rate as the number of kernels grow; 2) the training complexity is linear in the number of training examples; 3) very few iterations are enough to reach good solutions. Experiments on three standard benchmark databases support our claims.

Learning object categories from small samples is a challenging problem, where machine learning tools can in general provide very few guarantees. Exploiting prior knowledge may be useful to reproduce the human capability of recognizing objects even from only one single view. This paper presents an SVM-based model adaptation algorithm able to select and weight appropriately prior knowledge coming from different categories. The method relies on the solution of a convex optimization problem which ensures to have the minimal leave-one-out error on the training set. Experiments on a subset of the Caltech-256 database show that the proposed method produces better results than both choosing one single prior model, and transferring from all previous experience in a flat uninformative way.

One of the major challenges in developing autonomous systems is to make them able to recognize and categorize objects robustly. However, the appearance-based algorithms that are widely employed for robot perception do not explore the functionality of objects, described in terms of their affordances. These affordances (e.g., manipulation, grasping) are discriminative for object categories and are important cues for reliable robot performance in everyday environments. In this paper, we propose a strategy for object recognition that integrates both visual appearance and grasp affordance features. Following previous work, we hypothesize that additional grasp information improves object recognition, even if we reconstruct the grasp modality from visual features using a mapping function. We considered two different representations for the grasp modality: (1) motor information of the hand posture while grasping and (2) a more general grasp affordance descriptor. Using a multi-modal classifier we show that having real grasp information significantly boost object recognition. This improvement is preserved, although to a lesser extent, if the grasp modality is reconstructed using the mapping function.

Learning a category from few examples is a challenging task for vision algorithms, while psychological studies have shown that humans are able to generalise correctly even from a single instance (one-shot learning). The most accredited hypothesis is that humans are able to exploit prior knowledge when learning a new related category. This paper presents an SVM-based model adaptation algorithm able to perform knowledge transfer for a new category when very limited examples are available. Using a leave-one-out estimate of the weighted error-rate the algorithm automatically decides from where to transfer (on which known category to rely), how much to transfer (the degree of adaptation) and if it is worth transferring something at all. Moreover a weighted least-squares loss function takes optimally care of data unbalance between negative and positive examples. Experiments presented on two different object category databases show that the proposed method is able to exploit previous knowledge avoiding negative transfer. The overall classification performance is increased compared to what would be achieved by starting from scratch. Furthermore as the number of already learned categories grows, the algorithm is able to learn a new category from one sample with increasing precision, i.e. it is able to perform one-shot learning. © 2009. The copyright of this document resides with its authors.

The state-of-the-art in control of hand prosthetics is far from optimal. The main control interface is represented by surface electromyography (EMG): the activation potentials of the remnants of large muscles of the stump are used in a non-natural way to control one or, at best, two degrees-of-freedom. This has two drawbacks: first, the dexterity of the prosthesis is limited, leading to poor interaction with the environment; second, the patient undergoes a long training time. As more dexterous hand prostheses are put on the market, the need for a finer and more natural control arises. Machine learning can be employed to this end. A desired feature is that of providing a pre-trained model to the patient, so that a quicker and better interaction can be obtained. To this end we propose model adaptation with least-squares SVMs, a technique that allows the automatic tuning of the degree of adaptation. We test the effectiveness of the approach on a database of EMG signals gathered from human subjects. We show that, when pre-trained models are used, the number of training samples needed to reach a certain performance is reduced, and the overall performance is increased, compared to what would be achieved by starting from scratch.

To fulfill the dream of having autonomous robots at home, there is a need for spatial representations augmented with semantic concepts. Vision has emerged recently as the key modality to recognize semantic categories like places (office, corridor, kitchen, etc). A crucial aspect of these semantic place representations is that they change over time, due to the dynamism of the world. This calls for visual algorithms able to learn from experience while at the same time managing the continuous flow of incoming data. This paper addresses these issues by presenting an SVM-based algorithm able to (a) learn continuously from experience with a fast updating rule, and (b) control the memory growth via a random forgetting mechanism while at the same time preserving an accuracy comparable to that of the batch algorithm. We apply our method to two different scenarios where learning from experience plays an important role: (1) continuous learning of visual places under dynamic changes, and (2) knowledge transfer of visual concepts across robot platforms. For both scenarios, results confirm the effectiveness of our approach.

Given a corpus of news items consisting of images accompanied by text captions, we want to find out `whos doing what, i.e. associate names and action verbs in the captions to the face and body pose of the persons in the images. We present a joint model for simultaneously solving the image-caption correspondences and learning visual appearance models for the face and pose classes occurring in the corpus. These models can then be used to recognize people and actions in novel images without captions. We demonstrate experimentally that our jointface and pose model solves the correspondence problem better than earlier models covering only the face, and that it can perform recognition of new uncaptioned images.

It is of prime importance in everyday human life to cope with and respond appropriately to events that are not foreseen by prior experience. Machines to a large extent lack the ability to respond appropriately to such inputs. An important class of unexpected events is defined by incongruent combinations of inputs from different modalities and therefore multimodal information provides a crucial cue for the identification of such events, e.g., the sound of a voice is being heard while the person in the field-of-view does not move her lips. In the project DIRAC ("Detection and Identification of Rare Audio-visual Cues") we have been developing algorithmic approaches to the detection of such events, as well as an experimental hardware platform to test it. An audio-visual platform ("AWEAR" - audio-visual wearable device) has been constructed with the goal to help users with disabilities or a high cognitive load to deal with unexpected events. Key hardware components include stereo panoramic vision sensors and 6-channel worn-behind-the-ear (hearing aid) microphone arrays. Data have been recorded to study audio-visual tracking, a/v scene/object classification and a/v detection of incongruencies.

We present a discriminative online algorithm with a bounded memory growth, which is based on the kernel-based Perceptron. Generally, the required memory of the kernel-based Perceptron for storing the online hypothesis is not bounded. Previous work has been focused on discarding part of the instances in order to keep the memory bounded. In the proposed algorithm the instances are not discarded, but projected onto the space spanned by the previous online hypothesis. We derive a relative mistake bound and compare our algorithm both analytically and empirically to the state-of-the-art Forgetron algorithm (Dekel et al, 2007). The first variant of our algorithm, called Projectron, outperforms the Forgetron. The second variant, called Projectron++, outperforms even the Perceptron.

Integrating information coming from different sensors is a fundamental capability for autonomous robots. For complex tasks like topological localization, it would be desirable to use multiple cues, possibly from different modalities, so to achieve robust performance. This paper proposes a new method for integrating multiple cues. For each cue we train a large margin classifier which outputs a set of scores indicating the confidence of the decision. These scores are then used as input to a Support Vector Machine, that learns how to weight each cue, for each class, optimally during training. We call this algorithm SVM-based Discriminative Accumulation Scheme (SVM-DAS). We applied our method to the topological localization task, using vision and laser-based cues. Experimental results clearly show the value of our approach.

Localization and context interpretation are two key competences for mobile robot systems. Visual place recognition, as opposed to purely geometrical models, holds promise of higher flexibility and association of semantics to the model. Ideally, a place recognition algorithm should be robust to dynamic changes and it should perform consistently when recognizing a room (for instance a corridor) in different geographical locations. Also, it should be able to categorize places, a crucial capability for transfer of knowledge and continuous learning. In order to test the suitability of visual recognition algorithms for these tasks, this paper presents a new database, acquired in three different labs across Europe. It contains image sequences of several rooms under dynamic changes, acquired at the same time with a perspective and omnidirectional camera, mounted on a socket. We assess this new database with an appearance- based algorithm that combines local features with support vector machines through an ad-hoc kernel. Results show the effectiveness of the approach and the value of the database.

Vision-based place recognition is a desirable feature for an autonomous mobile system. In order to work in realistic scenarios, visual recognition algorithms should be adaptive, i.e. should be able to learn from experience and adapt continuously to changes in the environment. This paper presents a discriminative incremental learning approach to place recognition. We use a recently introduced version of the incremental SVM, which allows to control the memory requirements as the system updates its internal representation. At the same time, it preserves the recognition performance of the batch algorithm. In order to assess the method, we acquired a database capturing the intrinsic variability of places over time. Extensive experiments show the power and the potential of the approach.

A distinctive feature of intelligent systems is their capability to analyze their level of expertise for a given task; in other words, they know what they know. As a way towards this ambitious goal, this paper presents a recognition algorithm able to measure its own level of confidence and, in case of uncertainty, to seek for extra information so to increase its own knowledge and ultimately achieve better performance. We focus on the visual place recognition problem for topological localization, and we take an SVM approach. We propose a new method for measuring the confidence level of the classification output, based on the distance of a test image and the average distance of training vectors. This method is combined with a discriminative accumulation scheme for cue integration. We show with extensive experiments that the resulting algorithm achieves better performances for two visual cues than the classic single cue SVM on the same task, while minimising the computational load. More important, our method provides a reliable measure of the level of confidence of the decision.

An important competence for a mobile robot system is the ability to localize and perform context interpretation. This is required to perform basic navigation and to facilitate local specific services. Usually localization is performed based on a purely geometric model. Through use of vision and place recognition a number of opportunities open up in terms of flexibility and association of semantics to the model. To achieve this we present an appearance based method for place recognition. The method is based on a large margin classifier in combination with a rich global image descriptor. The method is robust to variations in illumination and minor scene changes. The method is evaluated across several different cameras, changes in time-of-day and weather conditions. The results clearly demonstrate the value of the approach.

An important competence for a mobile robot system is the ability to localize and perform context interpretation. This is required to perform basic navigation and to facilitate local specific services. Usually localization is performed based on a purely geometric model. Through use of vision and place recognition a number of opportunities open up in terms of flexibility and association of semantics to the model. To achieve this, the present paper presents an appearance based method for place recognition. The method is based on a large margin classifier in combination with a rich global image descriptor. The method is robust to variations in illumination and minor scene changes. The method is evaluated across several different cameras, changes in time-of-day and weather conditions. The results clearly demonstrate the value of the approach

Although a considerable amount of work has been published on material classification, relatively little of it studies situations with considerable variation within each class. Many experiments use the exact same sample, or different patches from the same image, for training and test sets. Thus, such studies are vulnerable to effectively recognising one particular sample of a material as opposed to the material category. In contrast, this paper places firm emphasis on the capability to generalise to previously unseen instances of materials. We adopt an appearance-based strategy, and conduct experiments on a new database which contains several samples of each of eleven material categories, imaged under a variety of pose, illumination and scale conditions. Together, these sources of intra-class variation provide a stern challenge indeed for recognition. Somewhat surprisingly, the difference in performance between various state-of-the-art texture descriptors proves rather small in this task. On the other hand, we clearly demonstrate that very significant gains can be achieved via different SVM-based classification techniques. Selecting appropriate kernel parameters proves crucial. This motivates a novel recognition scheme based on a decision tree. Each node contains an SVM to split one class from all others with a kernel parameter optimal for that particular node. Hence, each decision is made using a different, optimal, class-specific metric. Experiments show the superiority of this approach over several state-of-the-art classifiers

Category detection is a lively area of research. While categorization algorithms tend to agree in using local descriptors, they differ in the choice of the classifier, with some using generative models and others discriminative approaches. This paper presents a method for object category detection which integrates a generative model with a discriminative classifier. For each object category, we generate an appearance codebook, which becomes a common vocabulary for the generative and discriminative methods. Given a query image, the generative part of the algorithm finds a set of hypotheses and estimates their support in location and scale. Then, the discriminative part verifies each hypothesis on the same codebook activations. The new algorithm exploits the strengths of both original methods, minimizing their weaknesses. Experiments on several databases show that our new approach performs better than its building blocks taken separately. Moreover, experiments on two challenging multi-scale databases show that our new algorithm outperforms previously reported results

Recent developments in computer vision have shown that local features can provide efficient representations suitable for robust object recognition. Support vector machines have been established as powerful learning algorithms with good generalization capabilities. We combine these two approaches and propose a general kernel method for recognition with local features. We show that the proposed kernel satisfies the Mercer condition and that it is, suitable for many established local feature frameworks. Large-scale recognition results are presented on three different databases, which demonstrate that SVMs with the proposed kernel perform better than standard matching techniques on local features. In addition, experiments on noisy and occluded images show that local feature representations significantly outperform global approaches.

This paper presents a kernel method that allows to combine color and shape information for appearance-based object recognition. It doesn't require to define a new common representation, but use the power of kernels to combine different representations together in an effective manner. These results are achieved using results of statis(cid:173) tical mechanics of spin glasses combined with Markov random fields via kernel functions. Experiments show an increase in recognition rate up to 5.92% with respect to conventional strategies.