Publications by Federico Bolelli

Visual Autoregressive (VAR) models face a severe "Memory Wall" on edge devices due to large model size and substantial KV-cache requirements. In this work, we analyze the Infinity VAR family (2B and 8B) and propose a compression pipeline for deployment on constrained NVIDIA Jetson systems. We diagnose critical bottlenecks: activation outliers reaching 353x the median and channel-skewed cache variance. To address this, we propose a hybrid pipeline combining SVDQuant—to structurally decouple weight outliers—and Asymmetric Per-Channel KV8 quantization. Our approach reduces the Infinity-8B footprint by 64% (37.1GB →13.3GB), fitting it on the mid-range Orin NX with a 4.1x speedup over Flux.1-dev (W4A4), while achieving superior aesthetic alignment (ImageReward 1.13 vs 0.935). Crucially, we also unlock entry-level feasibility for the Infinity-2B, compressing it from 16.0 to 7.71 GB to enable deployment on the Orin Nano. These results establish a new efficiency standard for high-fidelity generative AI at the edge. The code is available at https://github.com/Henvezz95/deepcompressor.

Multimodal Large Language Models (MLLMs) have achieved impressive performance across a variety of vision–language tasks. However, their internal working mechanisms remain largely underexplored. In his work, we introduce FG-TRACER, a framework designed to analyze the information flow between visual and textual modalities in MLLMs in free-form generation. Notably, our numerically stabilized computational method enables the first systematic analysis of multimodal information flow in underexplored domains such as image captioning and chain-of-thought (CoT) reasoning. We apply FG-TRACER to two state-of-the-art MLLMs—LLaMA 3.2-Vision and LLaVA 1.5—across three vision–language benchmarks—TextVQA, COCO 2014, and ChartQA—and we conduct a word-level analysis of multimodal integration. Our findings uncover distinct patterns of multimodal fusion across models and tasks, demonstrating that fusion dynamics are both model- and task-dependent. Overall, FG-TRACER offers a robust methodology for probing the internal mechanisms of MLLMs in free-form settings, providing new insights into their multimodal reasoning strategies. Our source code is publicly available at https://anonymous.4open.science/r/FG-TRACER-CB5A/.

High-resolution histological imaging provides essential detail for quantitative brain modeling, yet acquiring whole-brain data at micrometer scale remains technically and economically challenging. This work introduces Brain-SR, a diffusion-based super-resolution framework designed to reconstruct high-resolution cortical sections from low-resolution BigBrain data. Building upon the InvSR paradigm, our method performs resolution enhancement in the latent space of a pretrained variational autoencoder, guided by a task-specific noise-predictor network. A key contribution is a frequency-domain supervision term that compares the magnitude spectra of predicted and target patches, enforcing spectral consistency while remaining robust to local misalignments. Quantitative evaluations demonstrate that Brain-SR achieves substantial improvements in LPIPS (-27%) and FID (-58%) compared to baseline diffusion Super-Resolution, while spectral analysis confirms accurate recovery of the frequency distribution. The resulting reconstructions preserve neuronal structures consistent with high-resolution references, offering a practical step toward large-scale, morphologically faithful brain histology reconstruction. The code is publicly available to support reproducibility: https://github.com/AImageLab-zip/Brain-SR.

Barcodes are a critical technology in industrial automation, logistics, and retail, enabling fast and reliable data capture. While deep learning has significantly improved barcode localization accuracy, most modern architectures remain too computationally demanding for real-time deployment on embedded systems without dedicated hardware acceleration. In this work, we present BaFaLo (Barcode Fast Localizer), an ultra-lightweight segmentation-based neural network for barcode localization. Our model is specifically optimized for real-time performance on low-power CPUs while maintaining high localization accuracy for both 1D and 2D barcodes. It features a two-branch architecture—comprising a local feature extractor and a global context module—and is tailored for low-resolution inputs to improve inference speed further. We benchmark BaFaLo against several lightweight architectures for object detection or segmentation, including YOLO Nano, Fast-SCNN, BiSeNet V2, and ContextNet, using the BarBeR dataset. BaFaLo achieves the fastest inference time among all deep-learning models tested, operating at 57.62ms per frame on a single CPU core of a Raspberry Pi 3B+. Despite its compact design, it achieves a decoding rate nearly equivalent to YOLO Nano for 1D barcodes and only 3.5 percentage points lower for 2D barcodes while being approximately nine times faster.

CNNs and Transformer-based architectures are recently dominating the field of 3D medical segmentation. While CNNs face limitations in the local receptive field, Transformers require significant memory and data, making them less suitable for analyzing large 3D medical volumes. Consequently, fully convolutional network models like U-Net are still leading the 3D segmentation scenario. Although efforts have been made to reduce the Transformers computational complexity, such optimized models still struggle with content-based reasoning. This paper examines Mamba, a Recurrent Neural Network (RNN) based on State Space Models (SSMs), which achieves linear complexity and has outperformed Transformers in long-sequence tasks. Specifically, we assess Mamba’s performance in 3D medical segmentation using three widely recognized and commonly employed datasets and propose architectural enhancements to improve its segmentation effectiveness by mitigating the primary shortcomings of existing Mamba-based solutions.

This paper provides a detailed description of how to install, set up, and use "BarBeR" (Barcode Benchmark Repository) to reproduce the results presented in the ICPR 2024 paper "BarBeR: A Barcode Benchmarking Repository". The paper details the tests available in the repository and how the configuration parameters affect and influence experimental results.

Since their invention in 1949, barcodes have remained the preferred method for automatic data capture, playing a crucial role in supply chain management. To detect a barcode in an image, multiple algorithms have been proposed in the literature, with a significant increase of interest in the topic since the rise of deep learning. However, research in the field suffers from many limitations, including the scarcity of public datasets and code implementations, which hampers the reproducibility and reliability of published results. For this reason, we developed "BarBeR" (Barcode Benchmark Repository), a benchmark designed for testing and comparing barcode detection algorithms. This benchmark includes the code implementation of various detection algorithms for barcodes, along with a suite of useful metrics. It offers a range of test setups and can be expanded to include any localization algorithm. In addition, we provide a large, annotated dataset of 8748 barcode images, combining multiple public barcode datasets with standardized annotation formats for both detection and segmentation tasks. Finally, we share the results obtained from running the benchmark on our dataset, offering valuable insights into the performance of different algorithms.

We introduce Bits2Bites, the first publicly available dataset for occlusal classification from intra-oral scans, comprising 200 paired upper and lower dental arches annotated across multiple clinically relevant dimensions (sagittal, vertical, transverse, and midline relationships). Leveraging this resource, we propose a multi-task learning benchmark that jointly predicts five occlusal traits from raw 3D point clouds using state-of-the-art point-based neural architectures. Our approach includes extensive ablation studies assessing the benefits of multi-task learning against single-task baselines, as well as the impact of automatically-predicted anatomical landmarks as input features. Results demonstrate the feasibility of directly inferring comprehensive occlusion information from unstructured 3D data, achieving promising performance across all tasks. Our entire dataset, code, and pretrained models are publicly released to foster further research in automated orthodontic diagnosis.

Whole Slide Images (WSIs) are crucial in histological diagnostics, providing high-resolution insights into cellular structures. In addition to challenges like the gigapixel scale of WSIs and the lack of pixel-level annotations, privacy restrictions further complicate their analysis. For instance, in a hospital network, different facilities need to collaborate on WSI analysis without the possibility of sharing sensitive patient data. A more practical and secure approach involves sharing models capable of continual adaptation to new data. However, without proper measures, catastrophic forgetting can occur. Traditional continual learning techniques rely on storing previous data, which violates privacy restrictions. To address this issue, this paper introduces Context Optimization Multiple Instance Learning (CooMIL), a rehearsal-free continual learning framework explicitly designed for WSI analysis. It employs a WSI-specific prompt learning procedure to adapt classification models across tasks, efficiently preventing catastrophic forgetting. Evaluated on four public WSI datasets from TCGA projects, our model significantly outperforms state-of-the-art methods within the WSI-based continual learning framework. The source code is available at https://github.com/FrancescaMiccolis/CooMIL.

Testicular ultrasound imaging is vital for assessing male infertility, with testicular inhomogeneity serving as a key biomarker. However, subjective interpretation and the scarcity of publicly available datasets pose challenges to automated classification. In this study, we explore supervised and unsupervised pretraining strategies using a ResNet-based architecture, supplemented by diffusion-based generative models to synthesize realistic ultrasound images. Our results demonstrate that pretraining significantly enhances classification performance compared to training from scratch, and synthetic data can effectively substitute real images in the pretraining process, alleviating data-sharing constraints. These methods offer promising advancements toward robust, clinically valuable automated analysis of male infertility. The source code is publicly available at https://github.com/AImageLab-zip/TesticulUS/.