Tatiana Tommasi

Our comprehension of video streams depicting human activities is naturally multifaceted: in just a few moments, we can grasp what is happening, identify the relevance and interactions of objects in the scene, and forecast what will happen soon, everything all at once. To endow autonomous systems with such a holistic perception, learning how to correlate concepts, abstract knowledge across diverse tasks, and leverage tasks synergies when learning novel skills is essential. A significant step in this direction is EgoPack, a unified framework for understanding human activities across diverse tasks with minimal overhead. EgoPack promotes information sharing and collaboration among downstream tasks, essential for efficiently learning new skills. In this paper, we introduce Hier-EgoPack, which advances EgoPack by enabling reasoning also across diverse temporal granularities, which expands its applicability to a broader range of downstream tasks. To achieve this, we propose a novel hierarchical architecture for temporal reasoning equipped with a GNN layer specifically designed to tackle the challenges of multi-granularity reasoning effectively. We evaluate our approach on multiple Ego4D benchmarks involving both clip-level and frame-level reasoning, demonstrating how our hierarchical unified architecture effectively solves these diverse tasks simultaneously.

The widespread availability of pre-trained vision models has enabled numerous deep learning applications through their transferable representations. However, their computational and storage costs often limit practical deployment. Pruning-at-Initialization has emerged as a promising approach to compress models before training, enabling efficient task-specific adaptation. While conventional wisdom suggests that effective pruning requires task-specific data, this creates a challenge when downstream tasks are unknown in advance. In this paper, we investigate how data influences the pruning of pre-trained vision models. Surprisingly, pruning on one task retains the model's zero-shot performance also on unseen tasks. Furthermore, fine-tuning these pruned models not only improves performance on original seen tasks but can recover held-out tasks' performance. We attribute this phenomenon to the favorable loss landscapes induced by extensive pre-training on large-scale datasets.

The recently introduced odd-one-out anomaly detection task involves identifying the odd-looking instances within a multi-object scene. This problem presents several challenges for modern deep learning models, demanding spatial reasoning across multiple views and relational reasoning to understand context and generalize across varying object categories and layouts. We argue that these challenges must be addressed with efficiency in mind. To this end, we propose a DINO-based model that reduces the number of parameters by one third and shortens training time by a factor of three compared to the current state-of-the-art, while maintaining competitive performance. Our experimental evaluation also introduces a Multimodal Large Language Model baseline, providing insights into its current limitations in structured visual reasoning tasks. The project page can be found at https://silviochito.github.io/EfficientOddOneOut/

Deep Neural Networks (DNNs) have revolutionized several fields, including safety- and mission-critical applications, such as autonomous driving and space exploration. However, recent studies have highlighted that transient hardware faults can corrupt the model's output, leading to high misprediction probabilities. Since traditional reliability strategies, based on modular hardware, software replications, or matrix multiplication checksum impose a high overhead, there is a pressing need for efficient and effective hardening solutions tailored for DNNs. In this paper we present several network design choices and a training procedure that increase the robustness of standard deep models and thoroughly evaluate these strategies with experimental analyses on vision classification tasks. We name DieHardNet the specialized DNN obtained by applying all our hardening techniques that combine knowledge from experimental hardware faults characterization and machine learning studies. We conduct extensive ablation studies to quantify the reliability gain of each hardening component in DieHardNet. We perform over 10,000 instruction-level fault injections to validate our approach and expose DieHardNet executed on GPUs to an accelerated neutron beam equivalent to more than 570,000 years of natural radiation. Our evaluation demonstrates that DieHardNet can reduce the critical error rate (i.e., errors that modify the inference) up to 100 times compared to the unprotected baseline model, without causing any increase in inference time.

Interacting with real-world cluttered scenes poses several challenges to robotic agents that need to understand complex spatial dependencies among the observed objects to determine optimal pick sequences or efficient object retrieval strategies. Existing solutions typically manage simplified scenarios and focus on predicting pairwise object relationships following an initial object detection phase, but often overlook the global context or struggle with handling redundant and missing object relations. In this work, we present a modern take on visual relational reasoning for grasp planning. We introduce D3GD, a novel testbed that includes bin picking scenes with up to 35 objects from 97 distinct categories. Additionally, we propose D3G, a new end-to-end transformer-based dependency graph generation model that simultaneously detects objects and produces an adjacency matrix representing their spatial relationships. Recognizing the limitations of standard metrics, we employ the Average Precision of Relationships for the first time to evaluate model performance, conducting an extensive experimental benchmark. The obtained results establish our approach as the new state-of-the-art for this task, laying the foundation for future research in robotic manipulation.

Shifting deep learning models from lab environments to real-world settings entails preparing them to handle unforeseen conditions, including the chance of encountering novel objects from classes that were not included in their training data. Such occurrences can pose serious threats in various applications. The task of Semantic Novelty detection has attracted significant attention in the last years mainly on 2D images, overlooking the complex 3D nature of the real-world. In this study, we address this gap by examining the geometric structures of objects within 3D point clouds to detect semantic novelty effectively. We advance the field by introducing 3D-SeND, a method that harnesses a large-scale pre-trained model to extract patch-based object representations directly from its intermediate feature representation. These patches are used to characterize each known class precisely. At inference, a normalcy score is obtained by assessing whether a test sample can be reconstructed predominantly from patches of a single known class or from multiple classes. We evaluate 3D-SeND on real-world point cloud samples when the reference known data are synthetic and demonstrate that it excels in both standard and few-shot scenarios. Thanks to its patch-based object representation, it is possible to visualize 3D-SeND’s predictions with a valuable explanation of the decision process. Moreover, the inherent training-free nature of 3D-SeND allows for its immediate application to a wide array of real-world tasks, offering a compelling advantage over approaches that require a task-specific learning phase. Our code is available at https://paolotron.github.io/3DSend.github.io.