Massimo Bertozzi

Anthony Dell'Eva, Marco Orsingher, Massimo Bertozzi

Generating dense point clouds from sparse raw data benefits downstream 3D understanding tasks, but existing models are limited to a fixed upsampling ratio or to a short range of integer values. In this paper, we present APU-SMOG, a Transformer-based model for Arbitrary Point cloud Upsampling (APU). The sparse input is firstly mapped to a Spherical Mixture of Gaussians (SMOG) distribution, from which an arbitrary number of points can be sampled. Then, these samples are fed as queries to the Transformer decoder, which maps them back to the target surface. Extensive qualitative and quantitative evaluations show that APU-SMOG outperforms state-of-the-art fixed-ratio methods, while effectively enabling upsampling with any scaling factor, including non-integer values, with a single trained model. The code is available at https://github.com/apusmog/apusmog/

Tomaso Fontanini, Claudio Ferrari, Giuseppe Lisanti et al.

In semantic image synthesis the state of the art is dominated by methods that use customized variants of the SPatially-Adaptive DE-normalization (SPADE) layers, which allow for good visual generation quality and editing versatility. By design, such layers learn pixel-wise modulation parameters to de-normalize the generator activations based on the semantic class each pixel belongs to. Thus, they tend to overlook global image statistics, ultimately leading to unconvincing local style editing and causing global inconsistencies such as color or illumination distribution shifts. Also, SPADE layers require the semantic segmentation mask for mapping styles in the generator, preventing shape manipulations without manual intervention. In response, we designed a novel architecture where cross-attention layers are used in place of SPADE for learning shape-style correlations and so conditioning the image generation process. Our model inherits the versatility of SPADE, at the same time obtaining state-of-the-art generation quality, as well as improved global and local style transfer. Code and models available at https://github.com/TFonta/CA2SIS.

Leonardo Rossi, Vittorio Bernuzzi, Tomaso Fontanini et al.

The ability to understand the surrounding scene is of paramount importance for Autonomous Vehicles (AVs). This paper presents a system capable to work in an online fashion, giving an immediate response to the arise of anomalies surrounding the AV, exploiting only the videos captured by a dash-mounted camera. Our architecture, called MOVAD, relies on two main modules: a Short-Term Memory Module to extract information related to the ongoing action, implemented by a Video Swin Transformer (VST), and a Long-Term Memory Module injected inside the classifier that considers also remote past information and action context thanks to the use of a Long-Short Term Memory (LSTM) network. The strengths of MOVAD are not only linked to its excellent performance, but also to its straightforward and modular architecture, trained in a end-to-end fashion with only RGB frames with as less assumptions as possible, which makes it easy to implement and play with. We evaluated the performance of our method on Detection of Traffic Anomaly (DoTA) dataset, a challenging collection of dash-mounted camera videos of accidents. After an extensive ablation study, MOVAD is able to reach an AUC score of 82.17\%, surpassing the current state-of-the-art by +2.87 AUC. Our code will be available on https://github.com/IMPLabUniPr/movad/tree/movad_vad

Tomaso Fontanini, Claudio Ferrari, Massimo Bertozzi et al.

Semantic image synthesis (SIS) refers to the problem of generating realistic imagery given a semantic segmentation mask that defines the spatial layout of object classes. Most of the approaches in the literature, other than the quality of the generated images, put effort in finding solutions to increase the generation diversity in terms of style i.e. texture. However, they all neglect a different feature, which is the possibility of manipulating the layout provided by the mask. Currently, the only way to do so is manually by means of graphical users interfaces. In this paper, we describe a network architecture to address the problem of automatically manipulating or generating the shape of object classes in semantic segmentation masks, with specific focus on human faces. Our proposed model allows embedding the mask class-wise into a latent space where each class embedding can be independently edited. Then, a bi-directional LSTM block and a convolutional decoder output a new, locally manipulated mask. We report quantitative and qualitative results on the CelebMask-HQ dataset, which show our model can both faithfully reconstruct and modify a segmentation mask at the class level. Also, we show our model can be put before a SIS generator, opening the way to a fully automatic generation control of both shape and texture. Code available at https://github.com/TFonta/Semantic-VAE.

Filippo Botti, Alex Ergasti, Leonardo Rossi et al.

The goal of style transfer is, given a content image and a style source, generating a new image preserving the content but with the artistic representation of the style source. Most of the state-of-the-art architectures use transformers or diffusion-based models to perform this task, despite the heavy computational burden that they require. In particular, transformers use self- and cross-attention layers which have large memory footprint, while diffusion models require high inference time. To overcome the above, this paper explores a novel design of Mamba, an emergent State-Space Model (SSM), called Mamba-ST, to perform style transfer. To do so, we adapt Mamba linear equation to simulate the behavior of cross-attention layers, which are able to combine two separate embeddings into a single output, but drastically reducing memory usage and time complexity. We modified the Mamba's inner equations so to accept inputs from, and combine, two separate data streams. To the best of our knowledge, this is the first attempt to adapt the equations of SSMs to a vision task like style transfer without requiring any other module like cross-attention or custom normalization layers. An extensive set of experiments demonstrates the superiority and efficiency of our method in performing style transfer compared to transformers and diffusion models. Results show improved quality in terms of both ArtFID and FID metrics. Code is available at https://github.com/FilippoBotti/MambaST.

Marco Orsingher, Paolo Zani, Paolo Medici et al.

In this paper, a complete pipeline for image-based 3D reconstruction of urban scenarios is proposed, based on PatchMatch Multi-View Stereo (MVS). Input images are firstly fed into an off-the-shelf visual SLAM system to extract camera poses and sparse keypoints, which are used to initialize PatchMatch optimization. Then, pixelwise depths and normals are iteratively computed in a multi-scale framework with a novel depth-normal consistency loss term and a global refinement algorithm to balance the inherently local nature of PatchMatch. Finally, a large-scale point cloud is generated by back-projecting multi-view consistent estimates in 3D. The proposed approach is carefully evaluated against both classical MVS algorithms and monocular depth networks on the KITTI dataset, showing state of the art performances.

Marco Orsingher, Paolo Zani, Paolo Medici et al.

We present a framework, called MVG-NeRF, that combines classical Multi-View Geometry algorithms and Neural Radiance Fields (NeRF) for image-based 3D reconstruction. NeRF has revolutionized the field of implicit 3D representations, mainly due to a differentiable volumetric rendering formulation that enables high-quality and geometry-aware novel view synthesis. However, the underlying geometry of the scene is not explicitly constrained during training, thus leading to noisy and incorrect results when extracting a mesh with marching cubes. To this end, we propose to leverage pixelwise depths and normals from a classical 3D reconstruction pipeline as geometric priors to guide NeRF optimization. Such priors are used as pseudo-ground truth during training in order to improve the quality of the estimated underlying surface. Moreover, each pixel is weighted by a confidence value based on the forward-backward reprojection error for additional robustness. Experimental results on real-world data demonstrate the effectiveness of this approach in obtaining clean 3D meshes from images, while maintaining competitive performances in novel view synthesis.

Marco Orsingher, Paolo Zani, Paolo Medici et al.

Image datasets have been steadily growing in size, harming the feasibility and efficiency of large-scale 3D reconstruction methods. In this paper, a novel approach for scaling Multi-View Stereo (MVS) algorithms up to arbitrarily large collections of images is proposed. Specifically, the problem of reconstructing the 3D model of an entire city is targeted, starting from a set of videos acquired by a moving vehicle equipped with several high-resolution cameras. Initially, the presented method exploits an approximately uniform distribution of poses and geometry and builds a set of overlapping clusters. Then, an Integer Linear Programming (ILP) problem is formulated for each cluster to select an optimal subset of views that guarantees both visibility and matchability. Finally, local point clouds for each cluster are separately computed and merged. Since clustering is independent from pairwise visibility information, the proposed algorithm runs faster than existing literature and allows for a massive parallelization. Extensive testing on urban data are discussed to show the effectiveness and the scalability of this approach.

Alex Ergasti, Tomaso Fontanini, Claudio Ferrari et al.

Text-based person search (TBPS) is a problem that gained significant interest within the research community. The task is that of retrieving one or more images of a specific individual based on a textual description. The multi-modal nature of the task requires learning representations that bridge text and image data within a shared latent space. Existing TBPS systems face two major challenges. One is defined as inter-identity noise that is due to the inherent vagueness and imprecision of text descriptions and it indicates how descriptions of visual attributes can be generally associated to different people; the other is the intra-identity variations, which are all those nuisances e.g. pose, illumination, that can alter the visual appearance of the same textual attributes for a given subject. To address these issues, this paper presents a novel TBPS architecture named MARS (Mae-Attribute-Relation-Sensitive), which enhances current state-of-the-art models by introducing two key components: a Visual Reconstruction Loss and an Attribute Loss. The former employs a Masked AutoEncoder trained to reconstruct randomly masked image patches with the aid of the textual description. In doing so the model is encouraged to learn more expressive representations and textual-visual relations in the latent space. The Attribute Loss, instead, balances the contribution of different types of attributes, defined as adjective-noun chunks of text. This loss ensures that every attribute is taken into consideration in the person retrieval process. Extensive experiments on three commonly used datasets, namely CUHK-PEDES, ICFG-PEDES, and RSTPReid, report performance improvements, with significant gains in the mean Average Precision (mAP) metric w.r.t. the current state of the art.

Leonardo Rossi, Vittorio Bernuzzi, Tomaso Fontanini et al.

Due to the limitations of current optical and sensor technologies and the high cost of updating them, the spectral and spatial resolution of satellites may not always meet desired requirements. For these reasons, Remote-Sensing Single-Image Super-Resolution (RS-SISR) techniques have gained significant interest. In this paper, we propose Swin2-MoSE model, an enhanced version of Swin2SR. Our model introduces MoE-SM, an enhanced Mixture-of-Experts (MoE) to replace the Feed-Forward inside all Transformer block. MoE-SM is designed with Smart-Merger, and new layer for merging the output of individual experts, and with a new way to split the work between experts, defining a new per-example strategy instead of the commonly used per-token one. Furthermore, we analyze how positional encodings interact with each other, demonstrating that per-channel bias and per-head bias can positively cooperate. Finally, we propose to use a combination of Normalized-Cross-Correlation (NCC) and Structural Similarity Index Measure (SSIM) losses, to avoid typical MSE loss limitations. Experimental results demonstrate that Swin2-MoSE outperforms any Swin derived models by up to 0.377 - 0.958 dB (PSNR) on task of 2x, 3x and 4x resolution-upscaling (Sen2Venus and OLI2MSI datasets). It also outperforms SOTA models by a good margin, proving to be competitive and with excellent potential, especially for complex tasks. Additionally, an analysis of computational costs is also performed. Finally, we show the efficacy of Swin2-MoSE, applying it to a semantic segmentation task (SeasoNet dataset). Code and pretrained are available on https://github.com/IMPLabUniPr/swin2-mose/tree/official_code

Alex Ergasti, Giuseppe Gabriele Tarollo, Filippo Botti et al.

Joint audio-video (AV) generation is still a significant challenge in generative AI, primarily due to three critical requirements: quality of the generated samples, seamless multimodal synchronization and temporal coherence, with audio tracks that match the visual data and vice versa, and limitless video duration. In this paper, we present $^R$-FLAV, a novel transformer-based architecture that addresses all the key challenges of AV generation. We explore three distinct cross modality interaction modules, with our lightweight temporal fusion module emerging as the most effective and computationally efficient approach for aligning audio and visual modalities. Our experimental results demonstrate that $^R$-FLAV outperforms existing state-of-the-art models in multimodal AV generation tasks. Our code and checkpoints are available at https://github.com/ErgastiAlex/R-FLAV.

Munsif Ali, Leonardo Rossi, Massimo Bertozzi

Continual learning (CL) in the context of Generative Adversarial Networks (GANs) remains a challenging problem, particularly when it comes to learn from a few-shot (FS) samples without catastrophic forgetting. Current most effective state-of-the-art (SOTA) methods, like LFS-GAN, introduce a non-negligible quantity of new weights at each training iteration, which would become significant when considering the long term. For this reason, this paper introduces \textcolor{red}{\textbf{\underline{c}}}ontinual few-sh\textcolor{red}{\textbf{\underline{o}}}t learning with \textcolor{red}{\textbf{\underline{lo}}}w-\textcolor{red}{\textbf{\underline{r}}}ank adaptation in GANs named CoLoR-GAN, a framework designed to handle both FS and CL together, leveraging low-rank tensors to efficiently adapt the model to target tasks while reducing even more the number of parameters required. Applying a vanilla LoRA implementation already permitted us to obtain pretty good results. In order to optimize even further the size of the adapters, we challenged LoRA limits introducing a LoRA in LoRA (LLoRA) technique for convolutional layers. Finally, aware of the criticality linked to the choice of the hyperparameters of LoRA, we provide an empirical study to easily find the best ones. We demonstrate the effectiveness of CoLoR-GAN through experiments on several benchmark CL and FS tasks and show that our model is efficient, reaching SOTA performance but with a number of resources enormously reduced. Source code is available on \href{https://github.com/munsifali11/CoLoR-GAN}{Github.

Alex Ergasti, Filippo Botti, Tomaso Fontanini et al.

Diffusion models have become the most popular approach for high-quality image generation, but their high computational cost still remains a significant challenge. To address this problem, we propose U-Shape Mamba (USM), a novel diffusion model that leverages Mamba-based layers within a U-Net-like hierarchical structure. By progressively reducing sequence length in the encoder and restoring it in the decoder through Mamba blocks, USM significantly lowers computational overhead while maintaining strong generative capabilities. Experimental results against Zigma, which is currently the most efficient Mamba-based diffusion model, demonstrate that USM achieves one-third the GFlops, requires less memory and is faster, while outperforming Zigma in image quality. Frechet Inception Distance (FID) is improved by 15.3, 0.84 and 2.7 points on AFHQ, CelebAHQ and COCO datasets, respectively. These findings highlight USM as a highly efficient and scalable solution for diffusion-based generative models, making high-quality image synthesis more accessible to the research community while reducing computational costs.

Munsif Ali, Leonardo Rossi, Massimo Bertozzi

Few-shot and continual learning face two well-known challenges in GANs: overfitting and catastrophic forgetting. Learning new tasks results in catastrophic forgetting in deep learning models. In the case of a few-shot setting, the model learns from a very limited number of samples (e.g. 10 samples), which can lead to overfitting and mode collapse. So, this paper proposes a Continual Few-shot Teacher-Student technique for the generative adversarial network (CFTS-GAN) that considers both challenges together. Our CFTS-GAN uses an adapter module as a student to learn a new task without affecting the previous knowledge. To make the student model efficient in learning new tasks, the knowledge from a teacher model is distilled to the student. In addition, the Cross-Domain Correspondence (CDC) loss is used by both teacher and student to promote diversity and to avoid mode collapse. Moreover, an effective strategy of freezing the discriminator is also utilized for enhancing performance. Qualitative and quantitative results demonstrate more diverse image synthesis and produce qualitative samples comparatively good to very stronger state-of-the-art models.

Marco Orsingher, Anthony Dell'Eva, Paolo Zani et al.

Neural Radiance Fields (NeRF) have recently emerged as a powerful method for image-based 3D reconstruction, but the lengthy per-scene optimization limits their practical usage, especially in resource-constrained settings. Existing approaches solve this issue by reducing the number of input views and regularizing the learned volumetric representation with either complex losses or additional inputs from other modalities. In this paper, we present KeyNeRF, a simple yet effective method for training NeRF in few-shot scenarios by focusing on key informative rays. Such rays are first selected at camera level by a view selection algorithm that promotes baseline diversity while guaranteeing scene coverage, then at pixel level by sampling from a probability distribution based on local image entropy. Our approach performs favorably against state-of-the-art methods, while requiring minimal changes to existing NeRF codebases.

Vittorio Bernuzzi, Leonardo Rossi, Tomaso Fontanini et al.

Self-supervised learning (SSL) has recently emerged as a key strategy for building foundation models in remote sensing, where the scarcity of annotated data limits the applicability of fully supervised approaches. In this work, we introduce WaveMAE, a masked autoencoding framework tailored for multispectral satellite imagery. Unlike conventional pixel-based reconstruction, WaveMAE leverages a multi-level Discrete Wavelet Transform (DWT) to disentangle frequency components and guide the encoder toward learning scale-aware high-frequency representations. We further propose a Geo-conditioned Positional Encoding (GPE), which incorporates geographical priors via Spherical Harmonics, encouraging embeddings that respect both semantic and geospatial structure. To ensure fairness in evaluation, all methods are pretrained on the same dataset (fMoW-S2) and systematically evaluated on the diverse downstream tasks of the PANGAEA benchmark, spanning semantic segmentation, regression, change detection, and multilabel classification. Extensive experiments demonstrate that WaveMAE achieves consistent improvements over prior state-of-the-art approaches, with substantial gains on segmentation and regression benchmarks. The effectiveness of WaveMAE pretraining is further demonstrated by showing that even a lightweight variant, containing only 26.4% of the parameters, achieves state-of-the-art performance. Our results establish WaveMAE as a strong and geographically informed foundation model for multispectral remote sensing imagery.

Filippo Botti, Alex Ergasti, Tomaso Fontanini et al.

Diffusion Models have become very popular for Semantic Image Synthesis (SIS) of human faces. Nevertheless, their training and inference is computationally expensive and their computational requirements are high due to the quadratic complexity of attention layers. In this paper, we propose a novel architecture called SISMA, based on the recently proposed Mamba. SISMA generates high quality samples by controlling their shape using a semantic mask at a reduced computational demand. We validated our approach through comprehensive experiments with CelebAMask-HQ, revealing that our architecture not only achieves a better FID score yet also operates at three times the speed of state-of-the-art architectures. This indicates that the proposed design is a viable, lightweight substitute to transformer-based models.

Alex Ergasti, Claudio Ferrari, Tomaso Fontanini et al.

Semantic Image Synthesis (SIS) is among the most popular and effective techniques in the field of face generation and editing, thanks to its good generation quality and the versatility is brings along. Recent works attempted to go beyond the standard GAN-based framework, and started to explore Diffusion Models (DMs) for this task as these stand out with respect to GANs in terms of both quality and diversity. On the other hand, DMs lack in fine-grained controllability and reproducibility. To address that, in this paper we propose a SIS framework based on a novel Latent Diffusion Model architecture for human face generation and editing that is both able to reproduce and manipulate a real reference image and generate diversity-driven results. The proposed system utilizes both SPADE normalization and cross-attention layers to merge shape and style information and, by doing so, allows for a precise control over each of the semantic parts of the human face. This was not possible with previous methods in the state of the art. Finally, we performed an extensive set of experiments to prove that our model surpasses current state of the art, both qualitatively and quantitatively.

Anthony Dell'Eva, Fabio Pizzati, Massimo Bertozzi et al.

Image-to-image (i2i) networks struggle to capture local changes because they do not affect the global scene structure. For example, translating from highway scenes to offroad, i2i networks easily focus on global color features but ignore obvious traits for humans like the absence of lane markings. In this paper, we leverage human knowledge about spatial domain characteristics which we refer to as 'local domains' and demonstrate its benefit for image-to-image translation. Relying on a simple geometrical guidance, we train a patch-based GAN on few source data and hallucinate a new unseen domain which subsequently eases transfer learning to target. We experiment on three tasks ranging from unstructured environments to adverse weather. Our comprehensive evaluation setting shows we are able to generate realistic translations, with minimal priors, and training only on a few images. Furthermore, when trained on our translations images we show that all tested proxy tasks are significantly improved, without ever seeing target domain at training.