Traditional semantic segmentation methods can recognize at test time only the classes that are present in the training set. This is a significant limitation, especially for semantic segmentation algorithms mounted on intelligent autonomous systems, deployed in realistic settings. Regardless of how many classes the system has seen at training time, it is inevitable that unexpected, unknown objects will appear at test time. The failure in identifying such anomalies may lead to incorrect, even dangerous behaviors of the autonomous agent equipped with such segmentation model when deployed in the real world. Current state of the art of anomaly segmentation uses generative models, exploiting their incapability to reconstruct patterns unseen during training. However, training these models is expensive, and their generated artifacts may create false anomalies. In this paper we take a different route and we propose to address anomaly segmentation through prototype learning. Our intuition is that anomalous pixels are those that are dissimilar to all class prototypes known by the model. We extract class prototypes from the training data in a lightweight manner using a cosine similarity-based classifier. Experiments on StreetHazards show that our approach achieves the new state of the art, with a significant margin over previous works, despite the reduced computational overhead. Code is available at https://github.com/DarioFontanel/PAnS.
Federated Learning (FL) deals with learning a central model (i.e. the server) in privacy-constrained scenarios, where data are stored on multiple devices (i.e. the clients). The central model has no direct access to the data, but only to the updates of the parameters computed locally by each client. This raises a problem, known as statistical heterogeneity, because the clients may have different data distributions (i.e. domains). This is only partly alleviated by clustering the clients. Clustering may reduce heterogeneity by identifying the domains, but it deprives each cluster model of the data and supervision of others. Here we propose a novel Cluster-driven Graph Federated Learning (FedCG). In FedCG, clustering serves to address statistical heterogeneity, while Graph Convolutional Networks (GCNs) enable sharing knowledge across them. FedCG: i) identifies the domains via an FL-compliant clustering and instantiates domain-specific modules (residual branches) for each domain; ii) connects the domain-specific modules through a GCN at training to learn the interactions among domains and share knowledge; and iii) learns to cluster unsupervised via teacher-student classifier-training iterations and to address novel unseen test domains via their domain soft-assignment scores. Thanks to the unique interplay of GCN over clusters, FedCG achieves the state-of-the-art on multiple FL benchmarks.
Being able to segment unseen classes not observed during training is an important technical challenge in deep learning, because of its potential to reduce the expensive annotation required for semantic segmentation. Prior zero-label semantic segmentation works approach this task by learning visual-semantic embeddings or generative models. However, they are prone to overfitting on the seen classes because there is no training signal for them. In this paper, we study the challenging generalized zero-label semantic segmentation task where the model has to segment both seen and unseen classes at test time. We assume that pixels of unseen classes could be present in the training images but without being annotated. Our idea is to capture the latent information on unseen classes by supervising the model with self-produced pseudo-labels for unlabeled pixels. We propose a consistency regularizer to filter out noisy pseudo-labels by taking the intersections of the pseudo-labels generated from different augmentations of the same image. Our framework generates pseudo-labels and then retrain the model with human-annotated and pseudo-labelled data. This procedure is repeated for several iterations. As a result, our approach achieves the new state-of-the-art on PascalVOC12 and COCO-stuff datasets in the challenging generalized zero-label semantic segmentation setting, surpassing other existing methods addressing this task with more complex strategies.
In this work, we present a new, algorithm for multi-domain learning. Given a pretrained architecture and a set of visual domains received sequentially, the goal of multi-domain learning is to produce a single model performing a task in all the domains together. Recent works showed how we can address this problem by masking the internal weights of a given original conv-net through learned binary variables. In this work, we provide a general formulation of binary mask based models for multi-domain learning by affine transformations of the original network parameters. Our formulation obtains significantly higher levels of adaptation to new domains, achieving performances comparable to domain-specific models while requiring slightly more than 1 bit per network parameter per additional domain. Experiments on two popular benchmarks showcase the power of our approach, achieving performances close to state-of-the-art methods on the Visual Decathlon Challenge.
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.
Event cameras are novel bio-inspired sensors, which asynchronously capture pixel-level intensity changes in the form of "events". The innovative way they acquire data presents several advantages over standard devices, especially in poor lighting and high-speed motion conditions. However, the novelty of these sensors results in the lack of a large amount of training data capable of fully unlocking their potential. The most common approach implemented by researchers to address this issue is to leverage simulated event data. Yet, this approach comes with an open research question: how well simulated data generalize to real data? To answer this, we propose to exploit, in the event-based context, recent Domain Adaptation (DA) advances in traditional computer vision, showing that DA techniques applied to event data help reduce the sim-to-real gap. To this purpose, we propose a novel architecture, which we call Multi-View DA4E (MV-DA4E), that better exploits the peculiarities of frame-based event representations while also promoting domain invariant characteristics in features. Through extensive experiments, we prove the effectiveness of DA methods and MV-DA4E on N-Caltech101. Moreover, we validate their soundness in a real-world scenario through a cross-domain analysis on the popular RGB-D Object Dataset (ROD), which we extended to the event modality (RGB-E).
Unsupervised Domain Adaptation (UDA) is a key field in visual recognition, as it enables robust performances across different visual domains. In the deep learning era, the performance of UDA methods has been driven by better losses and by improved network architectures, specifically the addition of auxiliary domain-alignment branches to pre-trained backbones. However, all the neural architectures proposed so far are hand-crafted, which might hinder further progress. The current copious offspring of Neural Architecture Search (NAS) only alleviates hand-crafting so far, as it requires labels for model selection, which are not available in UDA, and is usually applied to the whole architecture, while using pre-trained models is a strict requirement for high performance. No prior work has addressed these aspects in the context of NAS for UDA. Here we propose an Adversarial Branch Architecture Search (ABAS) for UDA, to learn the auxiliary branch network from data without handcrafting. Our main contribution include i. a novel data-driven ensemble approach for model selection, to circumvent the lack of target labels, and ii. a pipeline to automatically search for the best performing auxiliary branch. To the best of our knowledge, ABAS is the first NAS method for UDA to comply with a pre-trained backbone, a strict requirement for high performance. ABAS outputs both the optimal auxiliary branch and its trained parameters. When applied to two modern UDA techniques, DANN and ALDA, it improves performance on three standard CV datasets (Office31, Office-Home and PACS). In all cases, ABAS robustly finds the branch architectures which yield best performances. Code will be released.
This paper presents our work for the ninth edition of the Dialogue System Technology Challenge (DSTC9). Our solution addresses the track number four: Simulated Interactive MultiModal Conversations. The task consists in providing an algorithm able to simulate a shopping assistant that supports the user with his/her requests. We address the task of response retrieval, that is the task of retrieving the most appropriate agent response from a pool of response candidates. Our approach makes use of a neural architecture based on transformer with a multi-attentive structure that conditions the response of the agent on the request made by the user and on the product the user is referring to. Final experiments on the SIMMC Fashion Dataset show that our approach achieves the second best scores on all the retrieval metrics defined by the organizers. The source code is available at https://github.com/D2KLab/dstc9-SIMMC.
Reducing the amount of supervision required by neural networks is especially important in the context of semantic segmentation, where collecting dense pixel-level annotations is particularly expensive. In this paper, we address this problem from a new perspective: Incremental Few-Shot Segmentation. In particular, given a pretrained segmentation model and few images containing novel classes, our goal is to learn to segment novel classes while retaining the ability to segment previously seen ones. In this context, we discover, against all beliefs, that fine-tuning the whole architecture with these few images is not only meaningful, but also very effective. We show how the main problems of end-to-end training in this scenario are i) the drift of the batch-normalization statistics toward novel classes that we can fix with batch renormalization and ii) the forgetting of old classes, that we can fix with regularization strategies. We summarize our findings with five guidelines that together consistently lead to the state of the art on the COCO and Pascal-VOC 2012 datasets, with different number of images per class and even with multiple learning episodes.