Image segmentation is a fundamental component of vision-based agricultural robotics, enabling accurate fruit localization, disease detection, and automated harvesting. However, real-world strawberry fields present significant challenges due to irregular fruit morphology, dense foliage occlusions, variable ripeness, and strong illumination variability. Moreover, segmentation models trained on a single dataset often fail to generalize across domains, limiting their practical deployment. This paper presents a comprehensive benchmark of classical computer vision methods, convolutional neural networks, instance-based models, and transformer-based architectures across three heterogeneous public strawberry datasets: Db1 (instance segmentation), Db2 (lesion segmentation), and Db3 (semantic segmentation). A unified preprocessing and evaluation framework is adopted to ensure fair comparison using standard metrics, including Intersection-over-Union (IoU), Dice coefficient, Precision, and Recall. Extensive in-domain experiments demonstrate that deep learning models significantly outperform classical approaches, with U-Net and SegFormer achieving IoU values above 0.95 on Db1 and up to 0.83 on Db3. Cross-domain zero-shot evaluations reveal a substantial generalization gap, with U-Net suffering IoU drops of up to 100\%, while SegFormer consistently exhibits improved robustness and reduced cross-domain degradation across most transfer scenarios. To our knowledge, these results establish the first systematic multi-dataset benchmark for strawberry segmentation under domain shift, highlighting the importance of transformer-based architectures for robust agricultural perception and providing practical insights for real-world robotic deployment.

Graph neural networks (GNNs) have achieved strong performance on graph learning tasks, but their message-passing mechanism makes it difficult to capture long-range structural dependencies and may lead to over-smoothing in deeper architectures. Graph transformers offer a possible remedy, yet existing approaches typically rely on learnable attention or additional structural encodings, which increase the number of trainable parameters and computational cost while not always exploiting graph structure explicitly. Motivated by these limitations, we propose GraFix++, a graph transformer based on a fixed (non-learnable) multi-head structural attention mechanism derived from graph kernels. Multiple attention heads capture a range of structural similarities between substructures in the input graph, while a GNN is employed to improve the node features extraction. The resulting graph transformer showcases an excellent performance on standard graph classification benchmarks, matching or surpassing a wide range of alternative graph-based approaches. Furthermore, our model benefits from a reduced number of learnable parameters and competitive training runtime. In our experiments, we extensively evaluate the impact of various graph kernels, multiple attention heads, and GNN integration, demonstrating their collective contribution to the model’s superior performance.

In this paper we investigate a generative AI approach for interpolating 3D point clouds of sea surface elevation data. Unlike classic AI models, which aim to reconstruct a dense surface directly from scattered samples, we leverage denoising diffusion probabilistic models (DDPMs) to cast the interpolation as the problem of stochastically generating realistic surfaces loosely constrained to the available points. Such a process exploits a neural model trained to discriminate noise from the underlying wave elevation signal to iteratively and gradually remove noise from an initial random scalar field. Different random fields generate alternative reconstructions from the same set of sparse input points, all exhibiting realistic spectra and effectively reconstructing high-wavenumber content, overcoming classical deterministic methods. We observed that the standard deviation of the generated sea surface elevations (at each grid point) strongly correlates with the mean absolute error against the true surface elevation, providing an inherent uncertainty quantification metric. This correlation offers a dual benefit: it provides a reliable assessment of the model’s performance in regions with sparse data and a way to fuse temporally related reconstructions by weighting the dispersion relation constraint with the generated surface fields. Our experiments show that the proposed method offers an improvement with respect to traditional interpolation techniques, considering the fidelity to the ground truth surface, computational efficiency, and the ability to quantify uncertainty.

In recent years, deep learning researchers have been increasingly interested in developing architectures able to operate on data abstracted as graphs, i.e., Graph Neural Networks (GNNs). At the same time, there has been a surge in the number of commercial AI systems deployed for real-world applications. At their core, the majority of these systems are based on black-box deep learning models, such as GNNs, greatly limiting their accountability and trustworthiness. The idea underpinning this paper is to exploit the representational power of graph variational autoencoders to learn an embedding space where a “convolution” between local structures and latent vectors can take place. The key intuition is that this embedding space can then be used to decode the learned latent vectors into more interpretable latent structures. Our experiments validate the performance of our model against widely used alternatives on standard graph benchmarks, while also showing the ability to probe the model decisions by visualising the learned structural patterns.

Autoencoders, a type of unsupervised model, are capable of learning effective latent representations of data without supervision, only requiring the decoder to be able to reconstruct the original data-point from its latent representation obtained through the encoder. When dealing with structured data, graph-autoencoders are one of the few effective ways to obtain such latent representations of graphs. However, when dealing with large, diverse graph datasets, autoencoders struggle to adapt to varying structures, leading to suboptimal encoding. In our paper, we introduce a novel approach called Mixture of Variational Graph Autoencoders, which addresses this limitation by introducing a mixture of encoder/decoder models which provide multiple local and class-specific models that better adapt to different patches of the data-space. An exhaustive experimental evaluation shows that our approach greatly outperforms the state of the art in reconstruction precision (Code: https://github.com/gdl-unive/MVGAE ).

In this paper, we present GGSD, a novel graph generative model based on 1) the spectral decomposition of the graph Laplacian matrix and 2) a diffusion process. Specifically, we propose to use a denoising model to sample eigenvectors and eigenvalues from which we can reconstruct the graph Laplacian and adjacency matrix. Using the Laplacian spectrum allows us to naturally capture the structural characteristics of the graph and work directly in the node space while avoiding the quadratic complexity bottleneck that limits the applicability of other diffusion-based methods. This, in turn, is accomplished by truncating the spectrum, which, as we show in our experiments, results in a faster yet accurate generative process, and by designing a novel transformer-based architecture linear in the number of nodes. Our permutation invariant model can also handle node features by concatenating them to the eigenvectors of each node. An extensive set of experiments on both synthetic and real-world graphs demonstrates the strengths of our model against state-of-the-art alternatives.

In this paper we introduce GraFix, a novel graph transformer with fixed structural attention. Inspired by recent works 1) harnessing the link between (graph) kernels and the attention mechanism of transformers and 2) favouring simple fixed (non-learnable) attentive patterns over the standard attention mechanism, we propose to use graph kernels, specifically the WL kernel, to replace the learnable attention mechanism of a transformer with a fixed one capturing the structural similarity between substructures in the input graphs. The resulting graph transformer showcases an excellent performance on standard graph classification benchmarks, performing on-par with and in some instances outperforming a wide variety of alternative graph neural network and graph transformer-based approaches while at the same time benefiting from a reduced number of learnable parameters and learning runtime.

Graph neural networks are increasingly becoming the framework of choice for graph-based machine learning. In this paper, we propose a new graph neural network architecture that substitutes classical message passing with an analysis of the local distribution of node features. To this end, we extract the distribution of features in the egonet for each local neighbourhood and compare them against a set of learned label distributions by taking the histogram intersection kernel. The similarity information is then propagated to other nodes in the network, effectively creating a message passing-like mechanism where the message is determined by the ensemble of the features. We perform an ablation study to evaluate the network’s performance under different choices of its hyper-parameters. Finally, we test our model on standard graph classification and regression benchmarks, and we find that it outperforms widely used alternative approaches, including both graph kernels and graph neural networks.

Spectral signatures have been used with great success in computer vision to characterise the local and global topology of 3D meshes. In this paper, we propose to use two widely used spectral signatures, the Heat Kernel Signature and the Wave Kernel Signature, to create node embeddings able to capture local and global structural information for a given graph. For each node, we concatenate its structural embedding with the one-hot encoding vector of the node feature (if available) and we define a kernel between two input graphs in terms of the Wasserstein distance between the respective node embeddings. Experiments on standard graph classification benchmarks show that our kernel performs favourably when compared to widely used alternative kernels as well as graph neural networks.

Graph neural networks are increasingly becoming the framework of choice for graph-based machine learning. In this paper we propose a new graph neural network architecture based on the soft-alignment of the graph node features against sets of learned points. In each layer of the network the input node features are transformed by computing their similarity with respect to a set of learned features. The similarity information is then propagated to other nodes in the network, effectively creating a message passing-like mechanism where each node of the graph individually learns what is the optimal message to pass to its neighbours. We perform an ablation study to evaluate the performance of the network under different choices of its hyper-parameters. Finally, we test our model on standard graph-classification benchmarks and we find that it outperforms widely used alternative approaches, including both graph kernels and graph neural networks.

The convolution operator at the core of many modern neural architectures can effectively be seen as performing a dot product between an input matrix and a filter. While this is readily applicable to data such as images, which can be represented as regular grids in the Euclidean space, extending the convolution operator to work on graphs proves more challenging, due to their irregular structure. In this article, we propose to use graph kernels, i.e., kernel functions that compute an inner product on graphs, to extend the standard convolution operator to the graph domain. This allows us to define an entirely structural model that does not require computing the embedding of the input graph. Our architecture allows to plug-in any type of graph kernels and has the added benefit of providing some interpretability in terms of the structural masks that are learned during the training process, similar to what happens for convolutional masks in traditional convolutional neural networks (CNNs). We perform an extensive ablation study to investigate the model hyperparameters' impact and show that our model achieves competitive performance on standard graph classification and regression datasets.

High Dynamic Range (HDR) imaging techniques aim to increase the range of luminance values captured from a scene. The literature counts many approaches to get HDR images out of low-range camera sensors, however most of them rely on multiple acquisitions producing ghosting effects when moving objects are present. In this paper we propose a novel HDR reconstruction method exploiting stereo Polarimetric Filter Array (PFA) cameras to simultaneously capture the scene with different polarized filters, producing intensity attenuations that can be related to the light polarization state. An additional linear polarizer is mounted in front of one of the two cameras, raising the degree of polarization of rays captured by the sensor. This leads to a larger attenuation range between channels regardless the scene lighting condition. By merging the data acquired by the two cameras, we can compute the actual light attenuation observed by a pixel at each channel and derive an equivalent exposure time, producing a HDR picture from a single polarimetric shot. The proposed technique results comparable to classic HDR approaches using multiple exposures, with the advantage of being a one-shot method.

The Average Mixing Kernel Signature is a novel spectral signature for points on non-rigid three-dimensional shapes. It is based on a quantum exploration process of the shape surface, where the average transition probabilities between the points of the shape are summarised in the finite-time average mixing kernel. A band-filtered spectral analysis of this kernel then yields the AMKS. Crucially, we show that opting for a finite time-evolution allows the signature to account for a mixing of the Laplacian eigenspaces, similar to what is observed in the presence of noise, explaining the increased noise robustness of this signature when compared to alternative signatures. We perform an extensive experimental analysis of the AMKS under a wide range of problem scenarios, evaluating the performance of our descriptor under different sources of noise (vertex jitter and topological), shape representations (mesh and point clouds), as well as when only a partial view of the shape is available. Our experiments show that the AMKS consistently outperforms two of the most widely used spectral signatures, the Heat Kernel Signature and the Wave Kernel Signature, and suggest that the AMKS should be the signature of choice for various compute vision problems, including as input of deep convolutional architectures for shape analysis.

In this paper, we investigate the continual learning of Vision Transformers (ViT) for the challenging exemplar-free scenario, with special focus on how to efficiently distill the knowledge of its crucial self-attention mechanism (SAM). Our work takes an initial step towards a surgical investigation of SAM for designing coherent continual learning methods in ViTs. We first carry out an evaluation of established continual learning regularization techniques. We then examine the effect of regularization when applied to two key enablers of SAM: (a) the contextualized embedding layers, for their ability to capture well-scaled representations with respect to the values, and (b) the prescaled attention maps, for carrying value-independent global contextual information. We depict the perks of each distilling strategy on two image recognition benchmarks (CIFAR100 and ImageNet-32) – while (a) leads to a better overall accuracy, (b) helps enhance the rigidity by maintaining competitive performances. Furthermore, we identify the limitation imposed by the symmetric nature of regularization losses. To alleviate this, we propose an asymmetric variant and apply it to the pooled output distillation (POD) loss adapted for ViTs. Our experiments confirm that introducing asymmetry to POD boosts its plasticity while retaining stability across (a) and (b). Moreover, we acknowledge low forgetting measures for all the compared methods, indicating that ViTs might be naturally inclined continual learners.1

Polarisation Filter Array (PFA) cameras allow the analysis of light polarisation state in a simple and cost-effective manner. Such filter arrays work as the Bayer pattern for colour cameras, sharing similar advantages and drawbacks. Among the others, the raw image must be demosaiced considering the local variations of the PFA and the characteristics of the imaged scene. Non-linear effects, like the cross-talk among neighbouring pixels, are difficult to explicitly model and suggest the potential advantage of a data-driven learning approach. However, the PFA cannot be removed from the sensor, making it difficult to acquire the ground-truth polarization state for training. In this work we propose a novel CNN-based model which directly demosaics the raw camera image to a per-pixel Stokes vector. Our contribution is twofold. First, we propose a network architecture composed by a sequence of Mosaiced Convolutions operating coherently with the local arrangement of the different filters. Second, we introduce a new method, employing a consumer LCD screen, to effectively acquire real-world data for training. The process is designed to be invariant by monitor gamma and external lighting conditions. We extensively compared our method against algorithmic and learning-based demosaicing techniques, obtaining a consistently lower error especially in terms of polarisation angle.

The design of machines and algorithms capable of learning in a dynamically changing environment has become an increasingly topical problem with the increase of the size and heterogeneity of data available to learning systems. As a consequence, the key issue of Continual Learning has become that of addressing the stability-plasticity dilemma of connectionist systems, as they need to adapt their model without forgetting previously acquired knowledge. Within this context, rehearsal-based methods i.e., solutions in where the learner exploits memory to revisit past data, has proven to be very effective, leading to performance at the state-of-the-art. In our study, we propose an analysis of the memory quantity/quality trade-off adopting various data reduction approaches to increase the number of instances storable in memory. In particular, we investigate complex instance compression techniques such as deep encoders, but also trivial approaches such as image resizing and linear dimensionality reduction. Our findings suggest that the optimal trade-off is severely skewed toward instance quantity, where rehearsal approaches with several heavily compressed instances easily outperform state-of-the-art approaches with the same amount of memory at their disposal. Further, in high memory configurations, deep approaches extracting spatial structure combined with extreme resizing (of the order of 8×8 images) yield the best results, while in memory-constrained configurations where deep approaches cannot be used due to their memory requirement in training, Extreme Learning Machines (ELM) offer a clear advantage.

One of the most promising techniques for the analysis of Spatio-Temporal ocean wave fields is stereo vision. Indeed, the reconstruction accuracy and resolution typically outperform other approaches like radars, satellites, etc. However, it is computationally expensive so its application is typically restricted to the analysis of short pre-recorded sequences. What prevents such methodology from being truly real-time is the final 3D surface estimation from a scattered, non-equispaced point cloud. Recently, we studied a novel approach exploiting the temporal dependence of subsequent frames to iteratively update the wave spectrum over time. Albeit substantially faster, the unpredictable convergence time of the optimization involved still prevents its usage as a continuously running remote sensing infrastructure. In this work, we build upon the same idea, but investigating the feasibility of a fully data-driven Machine Learning (ML) approach. We designed a novel Convolutional Neural Network that learns how to produce an accurate surface from the scattered elevation data of three subsequent frames. The key idea is to embed the linear dispersion relation into the model itself to physically relate the sparse points observed at different times. Assuming that the scattered data are uniformly distributed in the spatial domain, this has the same effect of increasing the sample density of each single frame. Experiments demonstrate how the proposed technique, even if trained with purely synthetic data, can produce accurate and physically consistent surfaces at five frames per second on a modern PC.

The dimension of the space underlying real-world networks has been shown to strongly influence the networks structural properties, from the degree distribution to the way the networks respond to diffusion and percolation processes. In this paper we propose a way to estimate the dimension of the manifold underlying a network that is based on Weyl’s law, a mathematical result that describes the asymptotic behaviour of the eigenvalues of the graph Laplacian. For the case of manifold graphs, the dimension we estimate is equivalent to the fractal dimension of the network, a measure of structural self-similarity. Through an extensive set of experiments on both synthetic and real-world networks we show that our approach is able to correctly estimate the manifold dimension. We compare this with alternative methods to compute the fractal dimension and we show that our approach yields a better estimate on both synthetic and real-world examples.

We propose a new fast fully unsupervised method to discover semantic patterns. Our algorithm is able to hierarchically find visual categories and produce a segmentation mask where previous methods fail. Through the modeling of what is a visual pattern in an image, we introduce the notion of "semantic levels" and devise a conceptual framework along with measures and a dedicated benchmark dataset for future comparisons. Our algorithm is composed by two phases. A filtering phase, which selects semantical hotsposts by means of an accumulator space, then a clustering phase which propagates the semantic properties of the hotspots on a superpixels basis. We provide both qualitative and quantitative experimental validation, achieving optimal results in terms of robustness to noise and semantic consistency. We also made code and dataset publicly available.

Data augmentation is a widely adopted approach to solve the large-data requirements of modern deep learning techniques by generating new data instances from an existing dataset. While there is a huge literature and experience on augmentation for vectorial or image-based data, there is relatively little work on graph-based representations. This is largely due to complex, non-Euclidean structure of graphs, which limits our abilities to determine operations that do not modify the original semantic grouping. In this paper, we propose an alternative method for enlarging the graph set of graph neural network datasets by creating new graphs and keeping the properties of the originals. The proposal starts from the assumptions that the graphs compose a set of smaller motifs into larger structures. To this end, we extract modules by grouping nodes in an unsupervised way, and then swap similar modules between different graphs reconstructing the missing connectivity based on the original edge statistics and node similarity. We then test the performance of the proposed augmentation approach against state-of-the-art approaches, showing that on datasets, where the information is dominated by structure rather than node labels, we obtain a significant improvement with respect to alternatives.