Publications

Open-Vocabulary Segmentation (OVS) aims at segmenting images from free-form textual concepts without predefined training classes. While existing vision-language models such as CLIP can generate segmentation masks by leveraging coarse spatial information from Vision Transformers, they face challenges in spatial localization due to their global alignment of image and text features. Conversely, self-supervised visual models like DINO excel in fine-grained visual encoding but lack integration with language. To bridge this gap, we present Talk2DINO, a novel hybrid approach that combines the spatial accuracy of DINOv2 with the language understanding of CLIP. Our approach aligns the textual embeddings of CLIP to the patch-level features of DINOv2 through a learned mapping function without the need to fine-tune the underlying backbones. At training time, we exploit the attention maps of DINOv2 to selectively align local visual patches with textual embeddings. We show that the powerful semantic and localization abilities of Talk2DINO can enhance the segmentation process, resulting in more natural and less noisy segmentations, and that our approach can also effectively distinguish foreground objects from the background. Experimental results demonstrate that Talk2DINO achieves state-of-the-art performance across several unsupervised OVS benchmarks.

Recently, the field of 3D medical segmentation has been dominated by deep learning models employing Convolutional Neural Networks (CNNs) and Transformer-based architectures, each with its distinctive strengths and limitations. CNNs are constrained by a local receptive field, whereas Transformer are hindered by their substantial memory requirements as well as their data hunger, making them not ideal for processing 3D medical volumes at a fine-grained level. For these reasons, fully convolutional neural networks, as nnU-Net, still dominate the scene when segmenting medical structures in large 3D medical volumes. Despite numerous advancements toward developing transformer variants with subquadratic time and memory complexity, these models still fall short in content-based reasoning. A recent breakthrough is Mamba, a Recurrent Neural Network (RNN) based on State Space Models (SSMs), outperforming Transformers in many long-context tasks (million-length sequences) on famous natural language processing and genomic benchmarks while keeping a linear complexity. In this paper, we evaluate the effectiveness of Mamba-based architectures in comparison to state-of-the-art convolutional and Transformer-based models for 3D medical image segmentation across three well-established datasets: Synapse Abdomen, MSD BrainTumor, and ACDC. Additionally, we address the primary limitations of existing Mamba-based architectures by proposing alternative architectural designs, hence improving segmentation performances. The source code is publicly available to ensure reproducibility and facilitate further research: https://github.com/LucaLumetti/TamingMambas.

This paper addresses the challenges of deploying deep learning models, specifically Binary Neural Networks (BNNs), on resource-constrained embedded devices within the Internet of Things context. As deep learning continues to gain traction in IoT applications, the need for efficient models that can learn continuously from incremental data streams without requiring extensive computational resources has become more pressing. We propose a solution that integrates Continual Learning with BNNs, utilizing replay memory to prevent catastrophic forgetting. Our method focuses on quantized neural networks, introducing the quantization also for the backpropagation step, significantly reducing memory and computational requirements. Furthermore, we enhance the replay memory mechanism by storing intermediate feature maps (i.e. latent replay) with 1bit precision instead of raw data, enabling efficient memory usage. In addition to well-known benchmarks, we introduce the DL-Hazmat dataset, which consists of over 140k high-resolution grayscale images of 64 hazardous material symbols. Experimental results show a significant improvement in model accuracy and a substantial reduction in memory requirements, demonstrating the effectiveness of our method in enabling deep learning applications on embedded devices in real-world scenarios. Our work expands the application of Continual Learning and BNNs for efficient on-device training, offering a promising solution for IoT and other resource-constrained environments.

Attention guides our gaze to fixate the proper location of the scene and holds it in that location for the deserved amount of time given current processing demands, before shifting to the next one. As such, gaze deployment crucially is a temporal process. Existing computational models have made significant strides in predicting spatial aspects of observer's visual scanpaths (where to look), while often putting on the background the temporal facet of attention dynamics (when). In this paper we present TPP-Gaze, a novel and principled approach to model scanpath dynamics based on Neural Temporal Point Process (TPP), that jointly learns the temporal dynamics of fixations position and duration, integrating deep learning methodologies with point process theory. We conduct extensive experiments across five publicly available datasets. Our results show the overall superior performance of the proposed model compared to state-of-the-art approaches.

Multimodal Large Language Models (MLLMs) have recently emerged as a powerful framework for extending the capabilities of Large Language Models (LLMs) to reason over non-textual modalities. However, despite their success, understanding how they integrate visual and textual information remains an open challenge. Among them, LLaMA~3.2-Vision represents a significant milestone in the development of open-source MLLMs, offering a reproducible and efficient architecture that competes with leading proprietary models, such as Claude 3 Haiku and GPT-4o mini. Motivated by these characteristics, we conduct the first systematic analysis of the information flow between vision and language in LLaMA~3.2-Vision. We analyze three visual question answering (VQA) benchmarks, covering the tasks of VQA on natural images---using both open-ended and multiple-choice question formats---as well as document VQA. These tasks require diverse reasoning capabilities, making them well-suited to reveal distinct patterns in multimodal reasoning. Our analysis unveils a four-stage reasoning strategy: an initial semantic interpretation of the question, an early-to-mid-layer multimodal fusion, a task-specific reasoning stage guided by the resulting multimodal embedding, and a final answer prediction stage. Furthermore, we reveal that multimodal fusion is task-dependent: in complex settings such as document VQA, the model postpones cross-modal integration until semantic reasoning over the question has been established. Overall, our findings offer new insights into the internal dynamics of MLLMs and contribute to advancing the interpretability of vision-language architectures. Our source code is available at https://github.com/AImageLab/MLLMs-FlowTracker.

Trajectory forecasting is crucial for video surveillance analytics, as it enables the anticipation of future movements for a set of agents, e.g., basketball players engaged in intricate interactions with long-term intentions. Deep generative models offer a natural learning approach for trajectory forecasting, yet they encounter difficulties in achieving an optimal balance between sampling fidelity and diversity. We address this challenge by leveraging Vector Quantized Variational Autoencoders (VQ-VAEs), which utilize a discrete latent space to tackle the issue of posterior collapse. Specifically, we introduce an instance-based codebook that allows tailored latent representations for each example. In a nutshell, the rows of the codebook are dynamically adjusted to reflect contextual information (i.e., past motion patterns extracted from the observed trajectories). In this way, the discretization process gains flexibility, leading to improved reconstructions. Notably, instance-level dynamics are injected into the codebook through low-rank updates, which restrict the customization of the codebook to a lower dimension space. The resulting discrete space serves as the basis of the subsequent step, which regards the training of a diffusion-based predictive model. We show that such a two-fold framework, augmented with instance-level discretization, leads to accurate and diverse forecasts, yielding state-of-the-art performance on three established benchmarks.