Stefano Ghidoni

Alberto Bacchin, Leonardo Barcellona, Matteo Terreran et al.

Robotic waste sorting poses significant challenges in both perception and manipulation, given the extreme variability of objects that should be recognized on a cluttered conveyor belt. While deep learning has proven effective in solving complex tasks, the necessity for extensive data collection and labeling limits its applicability in real-world scenarios like waste sorting. To tackle this issue, we introduce a data augmentation method based on a novel GAN architecture called wasteGAN. The proposed method allows to increase the performance of semantic segmentation models, starting from a very limited bunch of labeled examples, such as few as 100. The key innovations of wasteGAN include a novel loss function, a novel activation function, and a larger generator block. Overall, such innovations helps the network to learn from limited number of examples and synthesize data that better mirrors real-world distributions. We then leverage the higher-quality segmentation masks predicted from models trained on the wasteGAN synthetic data to compute semantic-aware grasp poses, enabling a robotic arm to effectively recognizing contaminants and separating waste in a real-world scenario. Through comprehensive evaluation encompassing dataset-based assessments and real-world experiments, our methodology demonstrated promising potential for robotic waste sorting, yielding performance gains of up to 5.8\% in picking contaminants. The project page is available at https://github.com/bach05/wasteGAN.git

Alberto Bacchin, Davide Allegro, Stefano Ghidoni et al.

Out-of-Distribution (OOD) detection in computer vision is a crucial research area, with related benchmarks playing a vital role in assessing the generalizability of models and their applicability in real-world scenarios. However, existing OOD benchmarks in the literature suffer from two main limitations: (1) they often overlook semantic shift as a potential challenge, and (2) their scale is limited compared to the large datasets used to train modern models. To address these gaps, we introduce SOOD-ImageNet, a novel dataset comprising around 1.6M images across 56 classes, designed for common computer vision tasks such as image classification and semantic segmentation under OOD conditions, with a particular focus on the issue of semantic shift. We ensured the necessary scalability and quality by developing an innovative data engine that leverages the capabilities of modern vision-language models, complemented by accurate human checks. Through extensive training and evaluation of various models on SOOD-ImageNet, we showcase its potential to significantly advance OOD research in computer vision. The project page is available at https://github.com/bach05/SOODImageNet.git.

Laura Bragagnolo, Matteo Terreran, Davide Allegro et al.

Robust 3D human pose estimation is crucial to ensure safe and effective human-robot collaboration. Accurate human perception,however, is particularly challenging in these scenarios due to strong occlusions and limited camera viewpoints. Current 3D human pose estimation approaches are rather vulnerable in such conditions. In this work we present a novel approach for robust 3D human pose estimation in the context of human-robot collaboration. Instead of relying on noisy 2D features triangulation, we perform multi-view fusion on 3D skeletons provided by absolute monocular methods. Accurate 3D pose estimation is then obtained via reprojection error optimization, introducing limbs length symmetry constraints. We evaluate our approach on the public dataset Human3.6M and on a novel version Human3.6M-Occluded, derived adding synthetic occlusions on the camera views with the purpose of testing pose estimation algorithms under severe occlusions. We further validate our method on real human-robot collaboration workcells, in which we strongly surpass current 3D human pose estimation methods. Our approach outperforms state-of-the-art multi-view human pose estimation techniques and demonstrates superior capabilities in handling challenging scenarios with strong occlusions, representing a reliable and effective solution for real human-robot collaboration setups.

Laura Bragagnolo, Leonardo Barcellona, Stefano Ghidoni

Accurate 3D human pose estimation is fundamental for applications such as augmented reality and human-robot interaction. State-of-the-art multi-view methods learn to fuse predictions across views by training on large annotated datasets, leading to poor generalization when the test scenario differs. To overcome these limitations, we propose SkelSplat, a novel framework for multi-view 3D human pose estimation based on differentiable Gaussian rendering. Human pose is modeled as a skeleton of 3D Gaussians, one per joint, optimized via differentiable rendering to enable seamless fusion of arbitrary camera views without 3D ground-truth supervision. Since Gaussian Splatting was originally designed for dense scene reconstruction, we propose a novel one-hot encoding scheme that enables independent optimization of human joints. SkelSplat outperforms approaches that do not rely on 3D ground truth in Human3.6M and CMU, while reducing the cross-dataset error up to 47.8% compared to learning-based methods. Experiments on Human3.6M-Occ and Occlusion-Person demonstrate robustness to occlusions, without scenario-specific fine-tuning. Our project page is available here: https://skelsplat.github.io.

Ashkan Shakarami, Yousef Yeganeh, Azade Farshad et al.

This paper introduces Stress-Aware Learning, a resilient neural training paradigm in which deep neural networks dynamically adjust their optimization behavior - whether under stable training regimes or in settings with uncertain dynamics - based on the concept of Temporary (Elastic) and Permanent (Plastic) Deformation, inspired by structural fatigue in materials science. To instantiate this concept, we propose Plastic Deformation Optimizer, a stress-aware mechanism that injects adaptive noise into model parameters whenever an internal stress signal - reflecting stagnation in training loss and accuracy - indicates persistent optimization difficulty. This enables the model to escape sharp minima and converge toward flatter, more generalizable regions of the loss landscape. Experiments across six architectures, four optimizers, and seven vision benchmarks demonstrate improved robustness and generalization with minimal computational overhead. The code and 3D visuals will be available on GitHub: https://github.com/Stress-Aware-Learning/SAL.

Davide Allegro, Matteo Terreran, Stefano Ghidoni

In industrial scenarios, effective human-robot collaboration relies on multi-camera systems to robustly monitor human operators despite the occlusions that typically show up in a robotic workcell. In this scenario, precise localization of the person in the robot coordinate system is essential, making the hand-eye calibration of the camera network critical. This process presents significant challenges when high calibration accuracy should be achieved in short time to minimize production downtime, and when dealing with extensive camera networks used for monitoring wide areas, such as industrial robotic workcells. Our paper introduces an innovative and robust multi-camera hand-eye calibration method, designed to optimize each camera's pose relative to both the robot's base and to each other camera. This optimization integrates two types of key constraints: i) a single board-to-end-effector transformation, and ii) the relative camera-to-camera transformations. We demonstrate the superior performance of our method through comprehensive experiments employing the METRIC dataset and real-world data collected on industrial scenarios, showing notable advancements over state-of-the-art techniques even using less than 10 images. Additionally, we release an open-source version of our multi-camera hand-eye calibration algorithm at https://github.com/davidea97/Multi-Camera-Hand-Eye-Calibration.git.

Leonardo Barcellona, Andrii Zadaianchuk, Davide Allegro et al.

A world model provides an agent with a representation of its environment, enabling it to predict the causal consequences of its actions. Current world models typically cannot directly and explicitly imitate the actual environment in front of a robot, often resulting in unrealistic behaviors and hallucinations that make them unsuitable for real-world robotics applications. To overcome those challenges, we propose to rethink robot world models as learnable digital twins. We introduce DreMa, a new approach for constructing digital twins automatically using learned explicit representations of the real world and its dynamics, bridging the gap between traditional digital twins and world models. DreMa replicates the observed world and its structure by integrating Gaussian Splatting and physics simulators, allowing robots to imagine novel configurations of objects and to predict the future consequences of robot actions thanks to its compositionality. We leverage this capability to generate new data for imitation learning by applying equivariant transformations to a small set of demonstrations. Our evaluations across various settings demonstrate significant improvements in accuracy and robustness by incrementing actions and object distributions, reducing the data needed to learn a policy and improving the generalization of the agents. As a highlight, we show that a real Franka Emika Panda robot, powered by DreMa's imagination, can successfully learn novel physical tasks from just a single example per task variation (one-shot policy learning). Our project page can be found in: https://dreamtomanipulate.github.io/.

Davide Allegro, Shiyao Li, Stefano Ghidoni et al.

Current 3D mapping pipelines generally assume static environments, which limits their ability to accurately capture and reconstruct moving objects. To address this limitation, we introduce the novel task of active mapping of moving objects, in which a mapping agent must plan its trajectory while compensating for the object's motion. Our approach, Paparazzo, provides a learning-free solution that robustly predicts the target's trajectory and identifies the most informative viewpoints from which to observe it, to plan its own path. We also contribute a comprehensive benchmark designed for this new task. Through extensive experiments, we show that Paparazzo significantly improves 3D reconstruction completeness and accuracy compared to several strong baselines, marking an important step toward dynamic scene understanding. Project page: https://davidea97.github.io/paparazzo-page/

Leonardo Barcellona, Alberto Bacchin, Matteo Terreran et al.

The ability of a robot to pick an object, known as robot grasping, is crucial for several applications, such as assembly or sorting. In such tasks, selecting the right target to pick is as essential as inferring a correct configuration of the gripper. A common solution to this problem relies on semantic segmentation models, which often show poor generalization to unseen objects and require considerable time and massive data to be trained. To reduce the need for large datasets, some grasping pipelines exploit few-shot semantic segmentation models, which are capable of recognizing new classes given a few examples. However, this often comes at the cost of limited performance and fine-tuning is required to be effective in robot grasping scenarios. In this work, we propose to overcome all these limitations by combining the impressive generalization capability reached by foundation models with a high-performing few-shot classifier, working as a score function to select the segmentation that is closer to the support set. The proposed model is designed to be embedded in a grasp synthesis pipeline. The extensive experiments using one or five examples show that our novel approach overcomes existing performance limitations, improving the state of the art both in few-shot semantic segmentation on the Graspnet-1B (+10.5% mIoU) and Ocid-grasp (+1.6% AP) datasets, and real-world few-shot grasp synthesis (+21.7% grasp accuracy). The project page is available at: https://leobarcellona.github.io/showandgrasp.github.io/

Ashkan Shakarami, Yousef Yeganeh, Azade Farshad et al.

Activation functions are fundamental in deep neural networks and directly impact gradient flow, optimization stability, and generalization. Although ReLU remains standard because of its simplicity, it suffers from vanishing gradients and lacks adaptability. Alternatives like Swish and GELU introduce smooth transitions, but fail to dynamically adjust to input statistics. We propose VeLU, a Variance-enhanced Learning Unit as an activation function that dynamically scales based on input variance by integrating ArcTan-Sin transformations and Wasserstein-2 regularization, effectively mitigating covariate shifts and stabilizing optimization. Extensive experiments on ViT_B16, VGG19, ResNet50, DenseNet121, MobileNetV2, and EfficientNetB3 confirm VeLU's superiority over ReLU, ReLU6, Swish, and GELU on six vision benchmarks. The codes of VeLU are publicly available on GitHub.

Ashkan Shakarami, Azade Farshad, Yousef Yeganeh et al.

We propose UTS, a unit-based tissue segmentation framework for histopathology that classifies each fixed-size 32 * 32 tile, rather than each pixel, as the segmentation unit. This approach reduces annotation effort and improves computational efficiency without compromising accuracy. To implement this approach, we introduce a Multi-Level Vision Transformer (L-ViT), which benefits the multi-level feature representation to capture both fine-grained morphology and global tissue context. Trained to segment breast tissue into three categories (infiltrating tumor, non-neoplastic stroma, and fat), UTS supports clinically relevant tasks such as tumor-stroma quantification and surgical margin assessment. Evaluated on 386,371 tiles from 459 H&E-stained regions, it outperforms U-Net variants and transformer-based baselines. Code and Dataset will be available at GitHub.

Ashkan Shakarami, Lorenzo Nicole, Rocco Cappellesso et al.

Accurate and timely cancer diagnosis from histopathological slides is vital for effective clinical decision-making. This paper introduces DepViT-CAD, a deployable AI system for multi-class cancer diagnosis in histopathology. At its core is MAViT, a novel Multi-Attention Vision Transformer designed to capture fine-grained morphological patterns across diverse tumor types. MAViT was trained on expert-annotated patches from 1008 whole-slide images, covering 11 diagnostic categories, including 10 major cancers and non-tumor tissue. DepViT-CAD was validated on two independent cohorts: 275 WSIs from The Cancer Genome Atlas and 50 routine clinical cases from pathology labs, achieving diagnostic sensitivities of 94.11% and 92%, respectively. By combining state-of-the-art transformer architecture with large-scale real-world validation, DepViT-CAD offers a robust and scalable approach for AI-assisted cancer diagnostics. To support transparency and reproducibility, software and code will be made publicly available at GitHub.

Laura Bragagnolo, Matteo Terreran, Leonardo Barcellona et al.

Multi-human parsing is the task of segmenting human body parts while associating each part to the person it belongs to, combining instance-level and part-level information for fine-grained human understanding. In this work, we demonstrate that, while state-of-the-art approaches achieved notable results on public datasets, they struggle considerably in segmenting people with overlapping bodies. From the intuition that overlapping people may appear separated from a different point of view, we propose a novel training framework exploiting multi-view information to improve multi-human parsing models under occlusions. Our method integrates such knowledge during the training process, introducing a novel approach based on weak supervision on human instances and a multi-view consistency loss. Given the lack of suitable datasets in the literature, we propose a semi-automatic annotation strategy to generate human instance segmentation masks from multi-view RGB+D data and 3D human skeletons. The experiments demonstrate that the approach can achieve up to a 4.20\% relative improvement on human parsing over the baseline model in occlusion scenarios.

Loris Nanni, Stefano Ghidoni, Sheryl Brahnam

Features play a crucial role in computer vision. Initially designed to detect salient elements by means of handcrafted algorithms, features are now often learned by different layers in Convolutional Neural Networks (CNNs). This paper develops a generic computer vision system based on features extracted from trained CNNs. Multiple learned features are combined into a single structure to work on different image classification tasks. The proposed system was experimentally derived by testing several approaches for extracting features from the inner layers of CNNs and using them as inputs to SVMs that are then combined by sum rule. Dimensionality reduction techniques are used to reduce the high dimensionality of inner layers. The resulting vision system is shown to significantly boost the performance of standard CNNs across a large and diverse collection of image data sets. An ensemble of different topologies using the same approach obtains state-of-the-art results on a virus data set.

Loris Nanni, Alessandra Lumini, Stefano Ghidoni et al.

Classification of biological images is an important task with crucial application in many fields, such as cell phenotypes recognition, detection of cell organelles and histopathological classification, and it might help in early medical diagnosis, allowing automatic disease classification without the need of a human expert. In this paper we classify biomedical images using ensembles of neural networks. We create this ensemble using a ResNet50 architecture and modifying its activation layers by substituting ReLUs with other functions. We select our activations among the following ones: ReLU, leaky ReLU, Parametric ReLU, ELU, Adaptive Piecewice Linear Unit, S-Shaped ReLU, Swish , Mish, Mexican Linear Unit, Gaussian Linear Unit, Parametric Deformable Linear Unit, Soft Root Sign (SRS) and others. As a baseline, we used an ensemble of neural networks that only use ReLU activations. We tested our networks on several small and medium sized biomedical image datasets. Our results prove that our best ensemble obtains a better performance than the ones of the naive approaches. In order to encourage the reproducibility of this work, the MATLAB code of all the experiments will be shared at https://github.com/LorisNanni.

Andrea Franceschetti, Elisa Tosello, Nicola Castaman et al.

This paper proposes a detailed and extensive comparison of the Trust Region Policy Optimization and DeepQ-Network with Normalized Advantage Functions with respect to other state of the art algorithms, namely Deep Deterministic Policy Gradient and Vanilla Policy Gradient. Comparisons demonstrate that the former have better performances then the latter when asking robotic arms to accomplish manipulation tasks such as reaching a random target pose and pick &placing an object. Both simulated and real-world experiments are provided. Simulation lets us show the procedures that we adopted to precisely estimate the algorithms hyper-parameters and to correctly design good policies. Real-world experiments let show that our polices, if correctly trained on simulation, can be transferred and executed in a real environment with almost no changes.

Alessandro Malaguti, Marco Carraro, Mattia Guidolin et al.

This paper presents a novel real-time tracking system capable of improving body pose estimation algorithms in distributed camera networks. The first stage of our approach introduces a linear Kalman filter operating at the body joints level, used to fuse single-view body poses coming from different detection nodes of the network and to ensure temporal consistency between them. The second stage, instead, refines the Kalman filter estimates by fitting a hierarchical model of the human body having constrained link sizes in order to ensure the physical consistency of the tracking. The effectiveness of the proposed approach is demonstrated through a broad experimental validation, performed on a set of sequences whose ground truth references are generated by a commercial marker-based motion capture system. The obtained results show how the proposed system outperforms the considered state-of-the-art approaches, granting accurate and reliable estimates. Moreover, the developed methodology constrains neither the number of persons to track, nor the number, position, synchronization, frame-rate, and manufacturer of the RGB-D cameras used. Finally, the real-time performances of the system are of paramount importance for a large number of real-world applications.

Gianluca Maguolo, Loris Nanni, Stefano Ghidoni

Activation functions play a vital role in the training of Convolutional Neural Networks. For this reason, to develop efficient and performing functions is a crucial problem in the deep learning community. Key to these approaches is to permit a reliable parameter learning, avoiding vanishing gradient problems. The goal of this work is to propose an ensemble of Convolutional Neural Networks trained using several different activation functions. Moreover, a novel activation function is here proposed for the first time. Our aim is to improve the performance of Convolutional Neural Networks in small/medium size biomedical datasets. Our results clearly show that the proposed ensemble outperforms Convolutional Neural Networks trained with standard ReLU as activation function. The proposed ensemble outperforms with a p-value of 0.01 each tested stand-alone activation function; for reliable performance comparison we have tested our approach in more than 10 datasets, using two well-known Convolutional Neural Network: Vgg16 and ResNet50. MATLAB code used here will be available at https://github.com/LorisNanni.

Loris Nanni, Alessandra Lumini, Stefano Ghidoni

The aim of this work is to propose an ensemble of descriptors for Melanoma Classification, whose performance has been evaluated on validation and test datasets of the melanoma challenge 2018. The system proposed here achieves a strong discriminative power thanks to the combination of multiple descriptors. The proposed system represents a very simple yet effective way of boosting the performance of trained CNNs by composing multiple CNNs into an ensemble and combining scores by sum rule. Several types of ensembles are considered, with different CNN architectures along with different learning parameter sets. Moreover CNN are used as feature extractors: an input image is processed by a trained CNN and the response of a particular layer (usually the classification layer, but also internal layers can be employed) is treated as a descriptor for the image and used for training a set of Support Vector Machines (SVM).

Nicola Castaman, Elisa Tosello, Morris Antonello et al.

This paper proposes RUR53: an Unmanned Ground Vehicle able to autonomously navigate through, identify, and reach areas of interest; and there recognize, localize, and manipulate work tools to perform complex manipulation tasks. The proposed contribution includes a modular software architecture where each module solves specific sub-tasks and that can be easily enlarged to satisfy new requirements. Included indoor and outdoor tests demonstrate the capability of the proposed system to autonomously detect a target object (a panel) and precisely dock in front of it while avoiding obstacles. They show it can autonomously recognize and manipulate target work tools (i.e., wrenches and valve stems) to accomplish complex tasks (i.e., use a wrench to rotate a valve stem). A specific case study is described where the proposed modular architecture lets easy switch to a semi-teleoperated mode. The paper exhaustively describes description of both the hardware and software setup of RUR53, its performance when tests at the 2017 Mohamed Bin Zayed International Robotics Challenge, and the lessons we learned when participating at this competition, where we ranked third in the Gran Challenge in collaboration with the Czech Technical University in Prague, the University of Pennsylvania, and the University of Lincoln (UK).