We introduce PASTA (Perceptual Assessment System for explanaTion of Artificial Intelligence), a novel human-centric framework for evaluating eXplainable AI (XAI) techniques in computer vision. Our first contribution is the creation of the PASTA-dataset, the first large-scale benchmark that spans a diverse set of models and both saliency-based and concept-based explanation methods. This dataset enables robust, comparative analysis of XAI techniques based on human judgment. Our second contribution is an automated, data-driven benchmark that predicts human preferences using the PASTA-dataset. This scoring called PASTA-score method offers scalable, reliable, and consistent evaluation aligned with human perception. Additionally, our benchmark allows for comparisons between explanations across different modalities, an aspect previously unaddressed. We then propose to apply our scoring method to probe the interpretability of existing models and to build more human interpretable XAI methods.

This paper proposes a robust, reliable, and reproducible methodology for evaluating fairness in classification algorithms. Building upon established theoretical definitions of non-discrimination criteria—Independence, Separation, and Sufficiency—we present a comprehensive approach to quantifying fairness. Notably, our methodology extends beyond binary group comparisons to accommodate scenarios with multiple sensitive groups. Furthermore, we introduce a novel method for assessing fairness under the Sufficiency criterion, specifically focusing on Calibration. We assess the reliability and robustness of our model through applications to two real-world datasets.

Vision-language models (VLMs) often inherit the biases and unsafe associations present within their large-scale training dataset. While recent approaches mitigate unsafe behaviors, their evaluation focuses on how safe the model is on unsafe inputs, ignoring potential shortcomings on safe ones. In this paper, we first revise safety evaluation by introducing SafeGround, a new set of metrics that evaluate safety at different levels of granularity. With this metric, we uncover a surprising issue of training-based methods: they make the model less safe on safe inputs. From this finding, we take a different direction and explore whether it is possible to make a model safer without training, introducing Unsafe Weights Manipulation (UWM). UWM uses a calibration set of safe and unsafe instances to compare activations between safe and unsafe content, identifying the most important parameters for processing the latter. Their values are then manipulated via negation. Experiments show that UWM achieves the best tradeoff between safety and knowledge preservation, consistently improving VLMs on unsafe queries while outperforming even training-based state-of-the-art methods on safe ones.Warning: This paper includes unsafe and harmful content that may be disturbing. Such content has been blurred.

Abstract While Large Vision Language Models (LVLMs) demonstrate impressive capabilities, their substantial computational and memory requirements pose deployment challenges on resource-constrained edge devices. Current parameter reduction techniques primarily involve training LVLMs from small language models, but these methods offer limited flexibility and remain computationally intensive. We study a complementary route: compressing existing LVLMs by applying structured pruning to the language model backbone, followed by lightweight recovery training. Specifically, we investigate two structural pruning paradigms: layerwise and widthwise pruning, and pair them with supervised finetuning and knowledge distillation on logits and hidden states. Additionally, we assess the feasibility of conducting recovery training with only a small fraction of the available data. Our results show that widthwise pruning generally maintains better performance in low-resource scenarios, where computational resources are limited or there is insufficient finetuning data. As for the recovery training, finetuning only the multimodal projector is sufficient at small compression levels. Furthermore, a combination of supervised finetuning and hidden-state distillation yields optimal recovery across various pruning levels. Notably, effective recovery can be achieved using just 5% of the original data, while retaining over 95% of the original performance. Through empirical study on three representative LVLM families ranging from 3B to 7B parameters, this study offers actionable insights for practitioners to compress LVLMs without extensive computation resources or sufficient data.

Large vision-language models revolutionized image classification and semantic segmentation paradigms. However, they typically assume a pre-defined set of categories, or vocabulary, at test time for composing textual prompts. This assumption is impractical in scenarios with unknown or evolving semantic context. Here, we address this issue and introduce the Vocabulary-free Image Classification (VIC) task, which aims to assign a class from an unconstrained language-induced semantic space to an input image without needing a known vocabulary. VIC is challenging due to the vastness of the semantic space, which contains millions of concepts, including fine-grained categories. To address VIC, we propose Category Search from External Databases (CaSED), a training-free method that leverages a pre-trained vision-language model and an external database. CaSED first extracts the set of candidate categories from the most semantically similar captions in the database and then assigns the image to the best-matching candidate category according to the same vision-language model. Furthermore, we demonstrate that CaSED can be applied locally to generate a coarse segmentation mask that classifies image regions, introducing the task of Vocabulary-free Semantic Segmentation. CaSED and its variants outperform other more complex vision-language models, on classification and semantic segmentation benchmarks, while using much fewer parameters.

Vision-Language Models (VLMs) learn a shared feature space for text and images, enabling the comparison of inputs of different modalities. While prior works demonstrated that VLMs organize natural language representations into regular structures encoding composite meanings, it remains unclear if compositional patterns also emerge in the visual embedding space. In this work, we investigate compositionality in the image domain, where the analysis of compositional properties is challenged by noise and sparsity of visual data. We address these problems and propose a framework, called Geodesically Decomposable Embeddings (GDE), that approximates image representations with geometry-aware compositional structures in the latent space. We demonstrate that visual embeddings of pre-trained VLMs exhibit a compositional arrangement, and evaluate the effectiveness of this property in the tasks of compositional classification and group robustness. GDE achieves stronger performance in compositional classification compared to its counterpart method that assumes linear geometry of the latent space. Notably, it is particularly effective for group robustness, where we achieve higher results than task-specific solutions. Our results indicate that VLMs can automatically develop a human-like form of compositional reasoning in the visual domain, making their underlying processes more interpretable. Code is available at https://github.com/BerasiDavide/vlm_image_compositionality.

An old-school recipe for training a classifier is to (i) learn a good feature extractor and (ii) optimize a linear layer atop. When only a handful of samples are available per category, as in Few-Shot Adaptation (FSA), data are insufficient to fit a large number of parameters, rendering the above impractical. This is especially true with large pre-trained Vision-Language Models (VLMs), which motivated successful research at the intersection of Parameter-Efficient Fine-tuning (PEFT) and FSA. In this work, we start by analyzing the learning dynamics of PEFT techniques when trained on few-shot data from only a subset of categories, referred to as the "base" classes. We show that such dynamics naturally splits into two distinct phases: (i) task-level feature extraction and (ii) specialization to the available concepts. To accommodate this dynamic, we then depart from prompt- or adapter-based methods and tackle FSA differently. Specifically, given a fixed computational budget, we split it to (i) learn a task-specific feature extractor via PEFT and (ii) train a linear classifier on top. We call this scheme Two-Stage Few-Shot Adaptation (2SFS). Differently from established methods, our scheme enables a novel form of selective inference at a category level, i.e., at test time, only novel categories are embedded by the adapted text encoder, while embeddings of base categories are available within the classifier. Results with fixed hyperparameters across two settings, three backbones, and eleven datasets, show that 2SFS matches or surpasses the state-of-the-art, while established methods degrade significantly across settings. .

Attribute detection is crucial for many computer vision tasks, as it enables systems to describe properties such as color, texture, and material. Current approaches often rely on labor-intensive annotation processes which are inherently limited: objects can be described at an arbitrary level of detail (e.g., color vs. color shades), leading to ambiguities when the annotators are not instructed carefully. Furthermore, they operate within a predefined set of attributes, reducing scalability and adaptability to unforeseen downstream applications. We present Compositional Caching (ComCa), a training-free method for open-vocabulary attribute detection that overcomes these constraints. ComCa requires only the list of target attributes and objects as input, using them to populate an auxiliary cache of images by leveraging web-scale databases and Large Language Models to determine attribute-object compatibility. To account for the compositional nature of attributes, cache images receive soft attribute labels. Those are aggregated at inference time based on the similarity between the input and cache images, refining the predictions of underlying Vision-Language Models (VLMs). Importantly, our approach is model-agnostic, compatible with various VLMs. Experiments on public datasets demonstrate that ComCa significantly outperforms zero-shot and cache-based baselines, competing with recent training-based methods, proving that a carefully designed training-free approach can successfully address open-vocabulary attribute detection.

A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present Classifier-to-Bias (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection.

Recent video-language alignment models are trained on sets of videos, each with an associated positive caption and a negative caption generated by large language models. A problem with this procedure is that negative captions may introduce linguistic biases, i.e., concepts are seen only as negatives and never associated with a video. While a solution would be to collect videos for the negative captions, existing databases lack the fine-grained variations needed to cover all possible negatives. In this work, we study whether synthetic videos can help to overcome this issue. Our preliminary analysis with multiple generators shows that, while promising on some tasks, synthetic videos harm the performance of the model on others. We hypothesize this issue is linked to noise (semantic and visual) in the generated videos and develop a method, SynViTa, that accounts for those. SynViTa dynamically weights the contribution of each synthetic video based on how similar its target caption is w.r.t. the real counterpart. Moreover, a semantic consistency loss makes the model focus on fine-grained differences across captions, rather than differences in video appearance. Experiments show that, on average, SynViTa improves over existing methods on VideoCon test sets and SSv2-Temporal, SSv2-Events, and ATP-Hard benchmarks, being a first promising step for using synthetic videos when learning video-language models.

Traditional image classification requires a predefined list of semantic categories. In contrast, Large Multimodal Models (LMMs) can sidestep this requirement by classifying images directly using natural language (e.g., answering the prompt "What is the main object in the image?"). Despite this remarkable capability, most existing studies on LMM classification performance are surprisingly limited in scope, often assuming a closed-world setting with a predefined set of categories. In this work, we address this gap by thoroughly evaluating LMM classification performance in a truly open-world setting. We first formalize the task and introduce an evaluation protocol, defining various metrics to assess the alignment between predicted and ground truth classes. We then evaluate 13 models across 10 benchmarks, encompassing prototypical, non-prototypical, fine-grained, and very fine-grained classes, demonstrating the challenges LMMs face in this task. Further analyses based on the proposed metrics reveal the types of errors LMMs make, highlighting challenges related to granularity and fine-grained capabilities, showing how tailored prompting and reasoning can alleviate them.

Open-Vocabulary object detection models allow users to freely specify a class vocabulary in natural language at test time, guiding the detection of desired objects. However, vocabularies can be overly broad or even mis-specified, hampering the overall performance of the detector. In this work, we propose a plug-and-play Vocabulary Adapter (VocAda) to refine the user-defined vocabulary, automatically tailoring it to categories that are relevant for a given image. VocAda does not require any training, it operates at inference time in three steps: i) it uses an image captionner to describe visible objects, ii) it parses nouns from those captions, and iii) it selects relevant classes from the user-defined vocabulary, discarding irrelevant ones. Experiments on COCO and Objects365 with three state-of-the-art detectors show that VocAda consistently improves performance, proving its versatility. The code is open source.

Given a task and a set of steps composing it, Video Step Grounding (VSG) aims to detect which steps are performed in a video. Standard approaches for this task require a labeled training set (e.g., with step-level annotations or narrations), which may be costly to collect. Moreover, they process the full video offline, limiting their applications for scenarios requiring online decisions. Thus, in this work, we explore how to perform VSG online and without training. We achieve this by exploiting the zero-shot capabilities of recent Large Multimodal Models (LMMs). In particular, we use LMMs to predict the step associated with a restricted set of frames, without access to the whole video. We show that this online strategy without task-specific tuning outperforms offline and training-based models. Motivated by this finding, we develop Bayesian Grounding with Large Multimodal Models (BaGLM), further injecting knowledge of past frames into the LMM-based predictions. BaGLM exploits Bayesian filtering principles, modeling step transitions via (i) a dependency matrix extracted through large language models and (ii) an estimation of step progress. Experiments on three datasets show superior performance of BaGLM over state-of-the-art training-based offline methods.

Vision-Language Models (VLMs) have shown significant promise in Visual Question Answering (VQA) tasks by leveraging web-scale multimodal datasets. However, these models often struggle with continual learning due to catastrophic forgetting when adapting to new tasks. As an effective remedy to mitigate catastrophic forgetting, rehearsal strategy uses the data of past tasks upon learning new task. However, such strategy incurs the need of storing past data, which might not be feasible due to hardware constraints or privacy concerns. In this work, we propose the first data-free method that leverages the language generation capability of a VLM, instead of relying on external models, to produce pseudo-rehearsal data for addressing continual VQA. Our proposal, named as GaB, generates pseudo-rehearsal data by posing previous task questions on new task data. Yet, despite being effective, the distribution of generated questions skews towards the most frequently posed questions due to the limited and task-specific training data. To mitigate this issue, we introduce a pseudo-rehearsal balancing module that aligns the generated data towards the ground-truth data distribution using either the question meta-statistics or an unsupervised clustering method. We evaluate our proposed method on two recent benchmarks, i.e. VQACL- VQAv2 and CLOVE-function benchmarks. GaB outperforms all the data-free baselines with substantial improvement in maintaining VQA performance across evolving tasks, while being on-par with methods with access to the past data. Code and models are available at https://github.com/Deepayan137/GaB.

Supervised 3D part segmentation models are tailored for a fixed set of objects and parts, limiting their transferability to open-set, real-world scenarios. Recent works have explored vision-language models (VLMs) as a promising alternative, using multi-view rendering and textual prompting to identify object parts. However, naively applying VLMs in this context introduces several drawbacks, such as the need for meticulous prompt engineering, and fails to leverage the 3D geometric structure of objects. To address these limitations, we propose COPS, a COmprehensive model for Parts Segmentation that blends the semantics extracted from visual concepts and 3D geometry to effectively identify object parts. COPS renders a point cloud from multiple viewpoints, extracts 2D features, projects them back to 3D, and uses a novel geometric-aware feature aggregation procedure to ensure spatial and semantic consistency. Finally, it clusters points into parts and labels them. We demonstrate that COPS is efficient, scalable, and achieves zero-shot state-of-the-art performance across five datasets, covering synthetic and real-world data, texture-less and coloured objects, as well as rigid and non-rigid shapes. The code is available at https://3d-cops.github.io.

Recent progress in Text-to-Image (T2I) generative models has enabled high-quality image generation. As performance and accessibility increase, these models are gaining significant attraction and popularity: ensuring their fairness and safety is a priority to prevent the dissemination and perpetuation of biases. However, existing studies in bias detection focus on closed sets of predefined biases (e.g., gender, ethnicity). In this paper, we propose a general framework to identify, quantify, and explain biases in an open set setting, i.e. without requiring a predefined set. This pipeline leverages a Large Language Model (LLM) to propose biases starting from a set of captions. Next, these captions are used by the target generative model for generating a set of images. Finally, Vision Question Answering (VQA) is leveraged for bias evaluation. We show two variations of this framework: OpenBias and GradBias. OpenBias detects and quantifies biases, while GradBias determines the contribution of individual prompt words on biases. OpenBias effectively detects both well-known and novel biases related to people, objects, and animals and highly aligns with existing closed-set bias detection methods and human judgment. GradBias shows that neutral words can significantly influence biases and it outperforms several baselines, including state-of-the-art foundation models.

Text-to-image generative models are becoming increasingly popular and accessible to the general public. As these models see large-scale deployments, it is necessary to deeply investigate their safety and fairness to not disseminate and perpetuate any kind of biases. However, existing works focus on detecting closed sets of biases defined a priori, limiting the studies to well-known concepts. In this paper, we tackle the challenge of open-set bias detection in text-to-image generative models presenting OpenBias, a new pipeline that identifies and quantifies the severity of biases agnostically, without access to any precompiled set. OpenBias has three stages. In the first phase, we leverage a Large Language Model (LLM) to propose biases given a set of captions. Secondly, the target generative model produces images using the same set of captions. Lastly, a Vision Question Answering model recognizes the presence and extent of the previously proposed biases. We study the behavior of Stable Diffusion 1.5, 2, and XL emphasizing new biases, never investigated before. Via quantitative experiments, we demonstrate that OpenBias agrees with current closed-set bias detection methods and human judgement.

While excellent in transfer learning, Vision-Language models (VLMs) come with high computational costs due to their large number of parameters. To address this issue, removing parameters via model pruning is a viable solution. However, existing techniques for VLMs are task-specific, and thus require pruning the network from scratch for each new task of interest. In this work, we explore a new direction: Task-Agnostic Vision-Language Pruning (TA-VLP). Given a pretrained VLM, the goal is to find a unique pruned counterpart transferable to multiple unknown downstream tasks. In this challenging setting, the transferable representations already encoded in the pretrained model are a key aspect to preserve. Thus, we propose Multimodal Flow Pruning (MULTIFLOW), a first, gradient-free, pruning framework for TA-VLP where: (i) the importance of a parameter is expressed in terms of its magnitude and its information flow, by incorporating the saliency of the neurons it connects; and (ii) pruning is driven by the emergent (multimodal) distribution of the VLM parameters after pretraining. We benchmark eight state-of-the-art pruning algorithms in the context of TA-VLP, experimenting with two VLMs, three vision-language tasks, and three pruning ratios. Our experimental results show that MULTIFLOW outperforms recent sophisticated, combinatorial competitors in the vast majority of the cases, paving the way towards addressing TA- VLP. The code is publicly available at https://github.com/FarinaMatteo/multiflow.

Video anomaly detection (VAD) aims to temporally locate abnormal events in a video. Existing works mostly rely on training deep models to learn the distribution of normality with either video-level supervision, one-class supervision, or in an unsupervised setting. Training-based methods are prone to be domain-specific, thus being costly for practical deployment as any domain change will involve data collection and model training. In this paper, we radically depart from previous efforts and propose LAnguage-based VAD (LAVAD), a method tackling VAD in a novel, training-free paradigm, exploiting the capabilities of pre-trained large language models (LLMs) and existing vision-language models (VLMs). We leverage VLM-based captioning models to generate textual descriptions for each frame of any test video. With the textual scene description, we then devise a prompting mechanism to unlock the capability of LLMs in terms of temporal aggregation and anomaly score estimation, turning LLMs into an effective video anomaly detector. We further leverage modality-aligned VLMs and propose effective techniques based on cross-modal similarity for cleaning noisy captions and refining the LLM-based anomaly scores. We evaluate LAVAD on two large datasets featuring real-world surveillance scenarios (UCF-Crime and XD- Violence), showing that it outperforms both unsupervised and one-class methods without requiring any training or data collection.

Given an image and a target modification (e.g an image of the Eiffel tower and the text “without people and at night-time”), Compositional Image Retrieval (CIR) aims to retrieve the relevant target image in a database. While supervised approaches rely on annotating triplets that is costly (i.e. query image, textual modification, and target image), recent research sidesteps this need by using large-scale vision-language models (VLMs), performing Zero-Shot CIR (ZS-CIR). However, state-of-the-art approaches in ZS-CIR still require training task-specific, customized models over large amounts of image-text pairs. In this work, we proposeto tackle CIR in a training-free manner via our Compositional Image Retrieval through Vision-by-Language (CIReVL), a simple, yet human-understandable and scalable pipeline that effectively recombines large-scale VLMs with large language models (LLMs). By captioning the reference image using a pre-trained generative VLM and asking a LLM to recompose the caption based on the textual target modification for subsequent retrieval via e.g. CLIP, we achieve modular language reasoning. In four ZS-CIR benchmarks, we find competitive, in-part state-of-the-art performance - improving over supervised methods Moreover, the modularity of CIReVL offers simple scalability without re-training, allowing us to both investigate scaling laws and bottlenecks for ZS-CIR while easily scaling up to in parts more than double of previously reported results. Finally, we show that CIReVL makes CIR human-understandable by composing image and text in a modular fashion in the language domain, thereby making it intervenable, allowing to post-hoc re-align failure cases. Code will be released upon acceptance.

Vision-Language Models seamlessly discriminate among arbitrary semantic categories, yet they still suffer from poor generalization when presented with challenging examples. For this reason, Episodic Test-Time Adaptation (TTA) strategies have recently emerged as powerful techniques to adapt VLMs in the presence of a single unlabeled image. The recent literature on TTA is dominated by the paradigm of prompt tuning by Marginal Entropy Minimization, which, relying on online backpropagation, inevitably slows down inference while increasing memory. In this work, we theoretically investigate the properties of this approach and unveil that a surprisingly strong TTA method lies dormant and hidden within it. We term this approach ZERO (TTA with “zero” temperature), whose design is both incredibly effective and frustratingly simple: augment N times, predict, retain the most confident predictions, and marginalize after setting the Softmax temperature to zero. Remarkably, ZERO requires a single batched forward pass through the vision encoder only and no backward passes. We thoroughly evaluate our approach following the experimental protocol established in the literature and show that ZERO largely surpasses or compares favorably w.r.t. the state-of-the-art while being almost 10× faster and 13× more memory friendly than standard Test-Time Prompt Tuning. Thanks to its simplicity and comparatively negligible computation, ZERO can serve as a strong baseline for future work in this field. Code will be available.

Visual categories that largely share the same set of local parts cannot be discriminated based on part information alone, as they mostly differ in the way the local parts relate to the overall global structure of the object. We propose Relational Proxies, a novel approach that leverages the relational information between the global and local views of an object for encoding its semantic label, even for categories it has not encountered during training. Starting with a rigorous formalization of the notion of distinguishability between categories that share attributes, we prove the necessary and sufficient conditions that a model must satisfy in order to learn the underlying decision boundaries to tell them apart. We design Relational Proxies based on our theoretical findings and evaluate it on seven challenging fine-grained benchmark datasets and achieve state-of-the-art results on all of them, surpassing the performance of all existing works with a margin exceeding 4% in some cases. We additionally show that Relational Proxies also generalizes to the zero-shot setting, where it can efficiently leverage emergent relationships among attributes and image views to generalize to unseen categories, surpassing current state-of-the-art in both the non-generative and generative settings. Implementation is available at https://github.com/abhrac/relational-proxies.

State-of-the-art deep learning models are often trained with a large amount of costly labeled training data. However, requiring exhaustive manual annotations may degrade the model's generalizability in the limited-label regime.Semi-supervised learning and unsupervised learning offer promising paradigms to learn from an abundance of unlabeled visual data. Recent progress in these paradigms has indicated the strong benefits of leveraging unlabeled data to improve model generalization and provide better model initialization. In this survey, we review the recent advanced deep learning algorithms on semi-supervised learning (SSL) and unsupervised learning (UL) for visual recognition from a unified perspective. To offer a holistic understanding of the state-of-the-art in these areas, we propose a unified taxonomy. We categorize existing representative SSL and UL with comprehensive and insightful analysis to highlight their design rationales in different learning scenarios and applications in different computer vision tasks. Lastly, we discuss the emerging trends and open challenges in SSL and UL to shed light on future critical research directions.

Compositional Zero-Shot learning (CZSL) aims to recognize unseen compositions of state and object visual primitives seen during training. A problem with standard CZSL is the assumption of knowing which unseen compositions will be available at test time. In this work, we overcome this assumption operating on the open world setting, where no limit is imposed on the compositional space at test time, and the search space contains a large number of unseen compositions. To address this problem, we propose a new approach, Compositional Cosine Graph Embeddings (Co-CGE), based on two principles. First, Co-CGE models the dependency between states, objects and their compositions through a graph convolutional neural network. The graph propagates information from seen to unseen concepts, improving their representations. Second, since not all unseen compositions are equally feasible, and less feasible ones may damage the learned representations, Co-CGE estimates a feasibility score for each unseen composition, using the scores as margins in a cosine similarity-based loss and as weights in the adjacency matrix of the graphs. Experiments show that our approach achieves state-of-the-art performances in standard CZSL while outperforming previous methods in the open world scenario.

Humans are good at recomposing novel objects, i.e. they can identify commonalities between unknown objects from general structure to finer detail, an ability difficult to replicate by machines. We propose a framework, ISCO, to recompose an object using 3D superquadrics as semantic parts directly from 2D views without training a model that uses 3D supervision. To achieve this, we optimize the superquadric parameters that compose a specific instance of the object, comparing its rendered 3D view and 2D image silhouette. Our ISCO framework iteratively adds new superquadrics wherever the reconstruction error is high, abstracting first coarse regions and then finer details of the target object. With this simple coarse-to-fine inductive bias, ISCO provides consistent superquadrics for related object parts, despite not having any semantic supervision. Since ISCO does not train any neural network, it is also inherently robust to out-of-distribution objects. Experiments show that, compared to recent single instance superquadrics reconstruction approaches, ISCO provides consistently more accurate 3D reconstructions, even from images in the wild. Code available at https://github.com/ExplainableML/ISCO.

Zero-shot learning models achieve remarkable results on image classification for samples from classes that were not seen during training. However, such models must be trained from scratch with specialised methods: therefore, access to a training dataset is required when the need for zero-shot classification arises. In this paper, we aim to equip pre-trained models with zero-shot classification capabilities without the use of image data. We achieve this with our proposed Image-free Classifier Injection with Semantics (ICIS) that injects classifiers for new, unseen classes into pre-trained classification models in a post-hoc fashion without relying on image data. Instead, the existing classifier weights and simple class-wise descriptors, such as class names or attributes, are used. ICIS has two encoder-decoder networks that learn to reconstruct classifier weights from descriptors (and vice versa), exploiting (cross-)reconstruction and cosine losses to regularise the decoding process. Notably, ICIS can be cheaply trained and applied directly on top of pre-trained classification models. Experiments on benchmark ZSL datasets show that ICIS produces unseen classifier weights that achieve strong (generalised) zero-shot classification performance. Code is available at https://github.com/ExplainableML/ImageFreeZSL.

Fine-grained classification often requires recognizing specific object parts, such as beak shape and wing patterns for birds. Encouraging a fine-grained classification model to first detect such parts and then using them to infer the class could help us gauge whether the model is indeed looking at the right details better than with interpretability methods that provide a single attribution map. We propose PDiscoNet to discover object parts by using only image-level class labels along with priors encouraging the parts to be: discriminative, compact, distinct from each other, equivariant to rigid transforms, and active in at least some of the images. In addition to using the appropriate losses to encode these priors, we propose to use part-dropout, where full part feature vectors are dropped at once to prevent a single part from dominating in the classification, and part feature vector modulation, which makes the information coming from each part distinct from the perspective of the classifier. Our results on CUB, CelebA, and PartImageNet show that the proposed method provides substantially better part discovery performance than previous methods while not requiring any additional hyper-parameter tuning and without penalizing the classification performance. The code is available at https://github.com/robertdvdk/part_detection

Large-scale vision-language models (VLMs) like CLIP successfully find correspondences between images and text. Through the standard deterministic mapping process, an image or a text sample is mapped to a single vector in the embedding space. This is problematic: as multiple samples (images or text) can abstract the same concept in the physical world, deterministic embeddings do not reflect the inherent ambiguity in the embedding space. We propose ProbVLM, a probabilistic adapter that estimates probability distributions for the embeddings of pre-trained VLMs via inter/intra-modal alignment in a post-hoc manner without needing large-scale datasets or computing. On four challenging datasets, i.e., COCO, Flickr, CUB, and Oxford-flowers, we estimate the multi-modal embedding uncertainties for two VLMs, i.e., CLIP and BLIP, quantify the calibration of embedding uncertainties in retrieval tasks and show that ProbVLM outperforms other methods. Furthermore, we propose active learning and model selection as two real-world downstream tasks for VLMs and show that the estimated uncertainty aids both tasks. Lastly, we present a novel technique for visualizing the embedding distributions using a large-scale pre-trained latent diffusion model. Code is available at https://github.com/ExplainableML/ProbVLM

Recent advances in fine-grained representation learning leverage local-to-global (emergent) relationships for achieving state-of-the-art results. The relational representations relied upon by such methods, however, are abstract. We aim to deconstruct this abstraction by expressing them as interpretable graphs over image views. We begin by theoretically showing that abstract relational representations are nothing but a way of recovering transitive relationships among local views. Based on this, we design Transitivity Recovering Decompositions (TRD), a graph-space search algorithm that identifies interpretable equivalents of abstract emergent relationships at both instance and class levels, and with no post-hoc computations. We additionally show that TRD is provably robust to noisy views, with empirical evidence also supporting this finding. The latter allows TRD to perform at par or even better than the state-of-the-art, while being fully interpretable. Implementation is available at https://github.com/abhrac/trd.

Recent advances in large vision-language models have revolutionized the image classification paradigm. Despite showing impressive zero-shot capabilities, a pre-defined set of categories, a.k.a. the vocabulary, is assumed at test time for composing the textual prompts. However, such assumption can be impractical when the semantic context is unknown and evolving. We thus formalize a novel task, termed as Vocabulary-free Image Classification (VIC), where we aim to assign to an input image a class that resides in an unconstrained language-induced semantic space, without the prerequisite of a known vocabulary. VIC is a challenging task as the semantic space is extremely large, containing millions of concepts, with hard-to-discriminate fine-grained categories. In this work, we first empirically verify that representing this semantic space by means of an external vision-language database is the most effective way to obtain semantically relevant content for classifying the image. We then propose Category Search from External Databases (CaSED), a method that exploits a pre-trained vision-language model and an external vision-language database to address VIC in a training-free manner. CaSED first extracts a set of candidate categories from captions retrieved from the database based on their semantic similarity to the image, and then assigns to the image the best matching candidate category according to the same vision-language model. Experiments on benchmark datasets validate that CaSED outperforms other complex vision-language frameworks, while being efficient with much fewer parameters, paving the way for future research in this direction.

The goal of open-world compositional zero-shot learning (OW-CZSL) is to recognize compositions of state and objects in images, given only a subset of them during training and no prior on the unseen compositions. In this setting, models operate on a huge output space, containing all possible state-object compositions. While previous works tackle the problem by learning embeddings for the compositions jointly, here we revisit a simple CZSL baseline and predict the primitives, i.e. states and objects, independently. To ensure that the model develops primitive-specific features, we equip the state and object classifiers with separate, non-linear feature extractors. Moreover, we estimate the feasibility of each composition through external knowledge, using this prior to remove unfeasible compositions from the output space. Finally, we propose a new setting, i.e. CZSL under partial supervision (pCZSL), where either only objects or state labels are available during training, and we can use our prior to estimate the missing labels. Our model, Knowledge-Guided Simple Primitives (KG-SP), achieves state of the art in both OW-CZSL and pCZSL, surpassing most recent competitors even when coupled with semi-supervised learning techniques. Code available at: https://github.com/ExplainableML/KG-SP.

Fine-grained categories that largely share the same set of parts cannot be discriminated based on part information alone, as they mostly differ in the way the local parts relate to the overall global structure of the object. We propose Relational Proxies, a novel approach that leverages the relational information between the global and local views of an object for encoding its semantic label. Starting with a rigorous formalization of the notion of distinguishability between fine-grained categories, we prove the necessary and sufficient conditions that a model must satisfy in order to learn the underlying decision boundaries in the fine-grained setting. We design Relational Proxies based on our theoretical findings and evaluate it on seven challenging fine-grained benchmark datasets and achieve state-of-the-art results on all of them, surpassing the performance of all existing works with a margin exceeding 4% in some cases. We also experimentally validate our theory on fine-grained distinguishability and obtain consistent results across multiple benchmarks. Implementation is available at https://github.com/abhrac/relational-proxies.

Deep neural networks have enabled major progresses in semantic segmentation. However, even the most advanced neural architectures suffer from important limitations. First, they are vulnerable to catastrophic forgetting, i.e., they perform poorly when they are required to incrementally update their model as new classes are available. Second, they rely on large amount of pixel-level annotations to produce accurate segmentation maps. To tackle these issues, we introduce a novel incremental class learning approach for semantic segmentation taking into account a peculiar aspect of this task: since each training step provides annotation only for a subset of all possible classes, pixels of the background class exhibit a semantic shift. Therefore, we revisit the traditional distillation paradigm by designing novel loss terms which explicitly account for the background shift. Additionally, we introduce a novel strategy to initialize classifier's parameters at each step in order to prevent biased predictions toward the background class. Finally, we demonstrate that our approach can be extended to point- and scribble-based weakly supervised segmentation, modeling the partial annotations to create priors for unlabeled pixels. We demonstrate the effectiveness of our approach with an extensive evaluation on the Pascal-VOC, ADE20K, and Cityscapes datasets, significantly outperforming state-of-the-art methods.

Compositional Zero-Shot learning (CZSL) requires to recognize state-object compositions unseen during training. In this work, instead of assuming prior knowledge about the unseen compositions, we operate in the open world setting, where the search space includes a large number of unseen compositions some of which might be unfeasible. In this setting, we start from the cosine similarity between visual features and compositional embeddings. After estimating the feasibility score of each composition, we use these scores to either directly mask the output space or as a margin for the cosine similarity between visual features and compositional embeddings during training. Our experiments on two standard CZSL benchmarks show that all the methods suffer severe performance degradation when applied in the open world setting. While our simple CZSL model achieves state-of-the-art performances in the closed world scenario, our feasibility scores boost the performance of our approach in the open world setting, clearly outperforming the previous state of the art. Code is available at: https://github.com/ExplainableML/czsl.

Unsupervised Domain Adaptation (UDA) refers to the problem of learning a model in a target domain where labeled data are not available by leveraging information from annotated data in a source domain. Most deep UDA approaches operate in a single-source, single-target scenario, i.e., they assume that the source and the target samples arise from a single distribution. However, in practice most datasets can be regarded as mixtures of multiple domains. In these cases, exploiting traditional single-source, single-target methods for learning classification models may lead to poor results. Furthermore, it is often difficult to provide the domain labels for all data points, i.e. latent domains should be automatically discovered. This paper introduces a novel deep architecture which addresses the problem of UDA by automatically discovering latent domains in visual datasets and exploiting this information to learn robust target classifiers. Specifically, our architecture is based on two main components, i.e. a side branch that automatically computes the assignment of each sample to its latent domain and novel layers that exploit domain membership information to appropriately align the distribution of the CNN internal feature representations to a reference distribution. We evaluate our approach on publicly available benchmarks, showing that it outperforms state-of-the-art domain adaptation methods.

Despite their effectiveness in a wide range of tasks, deep architectures suffer from some important limitations. In particular, they are vulnerable to catastrophic forgetting, i.e. they perform poorly when they are required to update their model as new classes are available but the original training set is not retained. This paper addresses this problem in the context of semantic segmentation. Current strategies fail on this task because they do not consider a peculiar aspect of semantic segmentation: since each training step provides annotation only for a subset of all possible classes, pixels of the background class (i.e. pixels that do not belong to any other classes) exhibit a semantic distribution shift. In this work we revisit classical incremental learning methods, proposing a new distillation-based framework which explicitly accounts for this shift. Furthermore, we introduce a novel strategy to initialize classifier's parameters, thus preventing biased predictions toward the background class. We demonstrate the effectiveness of our approach with an extensive evaluation on the Pascal-VOC 2012 and ADE20K datasets, significantly outperforming state of the art incremental learning methods.

Recent unsupervised domain adaptation methods based on deep architectures have shown remarkable performance not only in traditional classification tasks but also in more complex problems involving structured predictions (e.g. semantic segmentation, depth estimation). Following this trend, in this paper we present a novel deep adaptation framework for estimating keypoints under domain shift, i.e, when the training (source) and the test (target) images significantly differ in terms of visual appearance. Our method seamlessly combines three different components: feature alignment, adversarial training and self-supervision. Specifically, our deep architecture leverages from domain-specific distribution alignment layers to perform target adaptation at the feature level. Furthermore, a novel loss is proposed which combines an adversarial term for ensuring aligned predictions in the output space and a geometric consistency term which guarantees coherent predictions between a target sample and its perturbed version. Our extensive experimental evaluation conducted on three publicly available benchmarks shows that our approach outperforms state-of-the-art domain adaptation methods in the 2D keypoint prediction task.

Motivated by recent advances in deep domain adaptation, this paper introduces a deep architecture for estimating 3D keypoints when the training (source) and the test (target) images greatly differ in terms of visual appearance (domain shift). Our approach operates by promoting domain distribution alignment in the feature space adopting batch normalization-based techniques. Furthermore, we propose to collect statistics about 3D keypoints positions of the source training data and to use this prior information to constrain predictions on the target domain introducing a loss derived from Multidimensional Scaling. We conduct an extensive experimental evaluation considering three publicly available benchmarks and show that our approach out-performs state-of-the-art domain adaptation methods for 3D keypoints predictions.

The ability to categorize is a cornerstone of visual intelligence, and a key functionality for artificial, autonomous visual machines. This problem will never be solved without algorithms able to adapt and generalize across visual domains. Within the context of domain adaptation and generalization, this paper focuses on the predictive domain adaptation scenario, namely the case where no target data are available and the system has to learn to generalize from annotated source images plus unlabeled samples with associated metadata from auxiliary domains. Our contribution is the first deep architecture that tackles predictive domain adaptation, able to leverage over the information brought by the auxiliary domains through a graph. Moreover, we present a simple yet effective strategy that allows us to take advantage of the incoming target data at test time, in a continuous domain adaptation scenario. Experiments on three benchmark databases support the value of our approach.

While today's robots are able to perform sophisticated tasks, they can only act on objects they have been trained to recognize. This is a severe limitation: any robot will inevitably see new objects in unconstrained settings, and thus will always have visual knowledge gaps. However, standard visual modules are usually built on a limited set of classes and are based on the strong prior that an object must belong to one of those classes. Identifying whether an instance does not belong to the set of known categories (i.e. open set recognition), only partially tackles this problem, as a truly autonomous agent should be able not only to detect what it does not know, but also to extend dynamically its knowledge about the world. We contribute to this challenge with a deep learning architecture that can dynamically update its known classes in an end-to-end fashion. The proposed deep network, based on a deep extension of a non-parametric model, detects whether a perceived object belongs to the set of categories known by the system and learns it without the need to retrain the whole system from scratch. Annotated images about the new category can be provided by an `oracle' (i.e. human supervision), or by autonomous mining of the Web. Experiments on two different databases and on a robot platform demonstrate the promise of our approach.

Deep networks have brought significant advances in robot perception, enabling to improve the capabilities of robots in several visual tasks, ranging from object detection and recognition to pose estimation, semantic scene segmentation and many others. Still, most approaches typically address visual tasks in isolation, resulting in overspecialized models which achieve strong performances in specific applications but work poorly in other (often related) tasks. This is clearly sub-optimal for a robot which is often required to perform simultaneously multiple visual recognition tasks in order to properly act and interact with the environment. This problem is exacerbated by the limited computational and memory resources typically available onboard to a robotic platform. The problem of learning flexible models which can handle multiple tasks in a lightweight manner has recently gained attention in the computer vision community and benchmarks supporting this research have been proposed. In this work we study this problem in the robot vision context, proposing a new benchmark, the RGB-D Triathlon, and evaluating state of the art algorithms in this novel challenging scenario. We also define a new evaluation protocol, better suited to the robot vision setting. Results shed light on the strengths and weaknesses of existing approaches and on open issues, suggesting directions for future research.

Current Domain Adaptation (DA) methods based on deep architectures assume that the source samples arise from a single distribution. However, in practice most datasets can be regarded as mixtures of multiple domains. In these cases exploiting single-source DA methods for learning target classifiers may lead to sub-optimal, if not poor, results. In addition, in many applications it is difficult to manually provide the domain labels for all source data points, i.e. latent domains should be automatically discovered. This paper introduces a novel Convolutional Neural Network (CNN) architecture which (i) automatically discovers latent domains in visual datasets and (ii) exploits this information to learn robust target classifiers. Our approach is based on the introduction of two main components, which can be embedded into any existing CNN architecture: (i) a side branch that automatically computes the assignment of a source sample to a latent domain and (ii) novel layers that exploit domain membership information to appropriately align the distribution of the CNN internal feature representations to a reference distribution. We test our approach on publicly-available datasets, showing that it outperforms state-of-the-art multi-source DA methods by a large margin.

A long standing problem in visual object categorization is the ability of algorithms to generalize across different testing conditions. The problem has been formalized as a covariate shift among the probability distributions generating the training data (source) and the test data (target) and several domain adaptation methods have been proposed to address this issue. While these approaches have considered the single source-single target scenario, it is plausible to have multiple sources and require adaptation to any possible target domain. This last scenario, named Domain Generalization (DG), is the focus of our work. Differently from previous DG methods which learn domain invariant representations from source data, we design a deep network with multiple domain-specific classifiers, each associated to a source domain. At test time we estimate the probabilities that a target sample belongs to each source domain and exploit them to optimally fuse the classifiers predictions. To further improve the generalization ability of our model, we also introduced a domain agnostic component supporting the final classifier. Experiments on two public benchmarks demonstrate the power of our approach.

Technological developments call for increasing perception and action capabilities of robots. Among other skills, vision systems that can adapt to any possible change in the working conditions are needed. Since these conditions are unpredictable, we need benchmarks which allow to assess the generalization and robustness capabilities of our visual recognition algorithms. In this work we focus on robotic kitting in unconstrained scenarios. As a first contribution, we present a new visual dataset for the kitting task. Differently from standard object recognition datasets, we provide images of the same objects acquired under various conditions where camera, illumination and background are changed. This novel dataset allows for testing the robustness of robot visual recognition algorithms to a series of different domain shifts both in isolation and unified. Our second contribution is a novel online adaptation algorithm for deep models, based on batch-normalization layers, which allows to continuously adapt a model to the current working conditions. Differently from standard domain adaptation algorithms, it does not require any image from the target domain at training time. We benchmark the performance of the algorithm on the proposed dataset, showing its capability to fill the gap between the performances of a standard architecture and its counterpart adapted offline to the given target domain.

Word embeddings are widely used in Natural Language Processing, mainly due to their success in capturing semantic information from massive corpora.However, their creation process does not allow the different meanings of a word to be automatically separated, as it conflates them into a single vector.We address this issue by proposing a new model which learns word and sense embeddings jointly.Our model exploits large corpora and knowledge from semantic networks in order to produce a unified vector space of word and sense embeddings.We evaluate the main features of our approach both qualitatively and quantitatively in a variety of tasks, highlighting the advantages of the proposed method in comparison to stateof-the-art word-and sense-based models.