The recent introduction of novel acquisition devices like the Leap Motion and Kinect allows to obtain a very informative description hand pose that can be exploited for accurate gesture recognition. This paper proposes recognition scheme explicitly targeted data. An ad-hoc feature set based on positions orientation fingertips is computed fed into multi-class SVM classifier in order recognize performed gestures. A features also extracted from depth combined with ones improve performance....

Deep learning architectures exhibit a critical drop of performance due to catastrophic forgetting when they are required incrementally learn new tasks. Contemporary incremental frameworks focus on image classification and object detection while in this work we formally introduce the problem for semantic segmentation which pixel-wise labeling is considered. To tackle task propose distill knowledge previous model retain information about previously learned classes, whilst updating current...

Microsoft Kinect had a key role in the development of consumer depth sensors being device that brought acquisition to mass market. Despite success this sensor, with introduction second generation, has completely changed technology behind sensor from structured light Time-Of-Flight. This paper presents comparison data provided by first and generation order explain achievements have been obtained switch technology. After an accurate analysis accuracy two under different conditions, sample...

Deep neural networks suffer from the major limitation of catastrophic forgetting old tasks when learning new ones. In this paper we focus on class incremental continual in semantic segmentation, where categories are made available over time while previous training data is not retained. The proposed scheme shapes latent space to reduce whilst improving recognition novel classes. Our framework driven by three components which also combine top existing techniques effortlessly. First, prototypes...

Deep networks allow to obtain outstanding results in semantic segmentation, however they need be trained a single shot with large amount of data. Continual learning settings where new classes are learned incremental steps and previous training data is no longer available challenging due the catastrophic forgetting phenomenon. Existing approaches typically fail when several performed or presence distribution shift background class. We tackle these issues by recreating for old outlining...

Accurate scene understanding from multiple sensors mounted on cars is a key requirement for autonomous driving systems. Nowadays, this task mainly performed through data-hungry deep learning techniques that need very large amounts of data to be trained. Due the high cost performing segmentation labeling, many synthetic datasets have been proposed. However, most them miss multi-sensor nature data, and do not capture significant changes introduced by variation daytime weather conditions. To...

Federated Learning (FL) has recently emerged as a possible way to tackle the domain shift in real-world Semantic Segmentation (SS) without compromising private nature of collected data. However, most existing works on FL unrealistically assume labeled data re-mote clients. Here we propose novel task (FFreeDA) which clients' is unlabeled and server accesses source dataset for pre-training only. To solve FFreeDA, LADD, leverages knowledge pre-trained model by employing self-supervision with...

Deep learning techniques have been widely used in autonomous driving systems for the semantic understanding of urban scenes. However, they need a huge amount labeled data training, which is difficult and expensive to acquire. A recently proposed workaround train deep networks using synthetic data, but domain shift between real world representations limits performance. In this work, novel Unsupervised Domain Adaptation (UDA) strategy introduced solve issue. The driven by three components:...

Deep learning frameworks allowed for a remarkable advancement in semantic segmentation, but the data hungry nature of convolutional networks has rapidly raised demand adaptation techniques able to transfer learned knowledge from label-abundant domains unlabeled ones. In this paper we propose an effective Unsupervised Domain Adaptation (UDA) strategy, based on feature clustering method that captures different modes distribution and groups features same class into tight well-separated...

The standard class-incremental continual learning setting assumes a set of tasks seen one after the other in fixed and pre-defined order. This is not very realistic federated environments where each client works independently an asynchronous manner getting data for different time-frames orders totally uncorrelated with ones. We introduce novel (AFCL) multiple happens at orderings time slots. tackle this task using prototype-based learning, representation loss, fractal pre-training, modified...

With the increasing availability of depth sensors, multimodal frameworks that combine color information with data are gaining interest. However, ground truth for semantic segmentation is burdensome to provide, thus making domain adaptation a significant research area. Yet most methods not able effectively handle data. Specifically, we address challenging source-free setting where performed without reusing source We propose MISFIT: MultImodal Source-Free Information fusion Transformer....

Time-of-Flight data is typically affected by a high level of noise and artifacts due to Multi-Path Interference (MPI). While various traditional approaches for ToF improvement have been proposed, machine learning techniques seldom applied this task, mostly the limited availability real world training with depth ground truth. In paper, we avoid rely on labeled in framework. A Coarse-Fine CNN, able exploit multi-frequency MPI correction, trained synthetic truth supervised way. parallel, an...

The semantic understanding of urban scenes is one the key components for an autonomous driving system. Complex deep neural networks this task require to be trained with a huge amount labeled data, which difficult and expensive acquire. A recently proposed workaround usage synthetic however differences between real world limit performance. We propose unsupervised domain adaptation strategy adapt supervised training data. learning exploits three components: standard on adversarial able exploit...

Federated Learning (FL) has recently emerged as a novel machine learning paradigm allowing to preserve privacy and account for the distributed nature of process in many real-world settings. Computer vision tasks deal with huge datasets often critical issues, therefore federated approaches have been presented exploit its privacy-preserving nature. Firstly, this paper introduces different FL settings used computer main challenges that need be tackled. Then, it provides comprehensive overview...

Open-vocabulary semantic segmentation enables models to identify novel object categories beyond their training data. While this flexibility represents a significant advancement, current approaches still rely on manually specified class names as input, creating an inherent bottleneck in real-world applications. This work proposes Vocabulary-Free Semantic Segmentation pipeline, eliminating the need for predefined vocabularies. Specifically, we address chicken-and-egg problem where users...