Alessio Tonioni

Mohamad Shahbazi, Evangelos Ntavelis, Alessio Tonioni et al. · eth-zurich

Pose-conditioned convolutional generative models struggle with high-quality 3D-consistent image generation from single-view datasets, due to their lack of sufficient 3D priors. Recently, the integration of Neural Radiance Fields (NeRFs) and generative models, such as Generative Adversarial Networks (GANs), has transformed 3D-aware generation from single-view images. NeRF-GANs exploit the strong inductive bias of neural 3D representations and volumetric rendering at the cost of higher computational complexity. This study aims at revisiting pose-conditioned 2D GANs for efficient 3D-aware generation at inference time by distilling 3D knowledge from pretrained NeRF-GANs. We propose a simple and effective method, based on re-using the well-disentangled latent space of a pre-trained NeRF-GAN in a pose-conditioned convolutional network to directly generate 3D-consistent images corresponding to the underlying 3D representations. Experiments on several datasets demonstrate that the proposed method obtains results comparable with volumetric rendering in terms of quality and 3D consistency while benefiting from the computational advantage of convolutional networks. The code will be available at: https://github.com/mshahbazi72/NeRF-GAN-Distillation

Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni et al.

The ability to generate highly realistic 2D images from mere text prompts has recently made huge progress in terms of speed and quality, thanks to the advent of image diffusion models. Naturally, the question arises if this can be also achieved in the generation of 3D content from such text prompts. To this end, a new line of methods recently emerged trying to harness diffusion models, trained on 2D images, for supervision of 3D model generation using view dependent prompts. While achieving impressive results, these methods, however, have two major drawbacks. First, rather than commonly used 3D meshes, they instead generate neural radiance fields (NeRFs), making them impractical for most real applications. Second, these approaches tend to produce over-saturated models, giving the output a cartoonish looking effect. Therefore, in this work we propose a novel method for generation of highly realistic-looking 3D meshes. To this end, we extend NeRF to employ an SDF backbone, leading to improved 3D mesh extraction. In addition, we propose a novel way to finetune the mesh texture, removing the effect of high saturation and improving the details of the output 3D mesh.

Fabio Tosi, Alessio Tonioni, Daniele De Gregorio et al.

We introduce a novel framework for training deep stereo networks effortlessly and without any ground-truth. By leveraging state-of-the-art neural rendering solutions, we generate stereo training data from image sequences collected with a single handheld camera. On top of them, a NeRF-supervised training procedure is carried out, from which we exploit rendered stereo triplets to compensate for occlusions and depth maps as proxy labels. This results in stereo networks capable of predicting sharp and detailed disparity maps. Experimental results show that models trained under this regime yield a 30-40% improvement over existing self-supervised methods on the challenging Middlebury dataset, filling the gap to supervised models and, most times, outperforming them at zero-shot generalization.

Mauro Comi, Yijiong Lin, Alex Church et al.

Humans rely on their visual and tactile senses to develop a comprehensive 3D understanding of their physical environment. Recently, there has been a growing interest in exploring and manipulating objects using data-driven approaches that utilise high-resolution vision-based tactile sensors. However, 3D shape reconstruction using tactile sensing has lagged behind visual shape reconstruction because of limitations in existing techniques, including the inability to generalise over unseen shapes, the absence of real-world testing, and limited expressive capacity imposed by discrete representations. To address these challenges, we propose TouchSDF, a Deep Learning approach for tactile 3D shape reconstruction that leverages the rich information provided by a vision-based tactile sensor and the expressivity of the implicit neural representation DeepSDF. Our technique consists of two components: (1) a Convolutional Neural Network that maps tactile images into local meshes representing the surface at the touch location, and (2) an implicit neural function that predicts a signed distance function to extract the desired 3D shape. This combination allows TouchSDF to reconstruct smooth and continuous 3D shapes from tactile inputs in simulation and real-world settings, opening up research avenues for robust 3D-aware representations and improved multimodal perception in robotics. Code and supplementary material are available at: https://touchsdf.github.io/

Pierluigi Zama Ramirez, Adriano Cardace, Luca De Luigi et al.

Availability of labelled data is the major obstacle to the deployment of deep learning algorithms for computer vision tasks in new domains. The fact that many frameworks adopted to solve different tasks share the same architecture suggests that there should be a way of reusing the knowledge learned in a specific setting to solve novel tasks with limited or no additional supervision. In this work, we first show that such knowledge can be shared across tasks by learning a mapping between task-specific deep features in a given domain. Then, we show that this mapping function, implemented by a neural network, is able to generalize to novel unseen domains. Besides, we propose a set of strategies to constrain the learned feature spaces, to ease learning and increase the generalization capability of the mapping network, thereby considerably improving the final performance of our framework. Our proposal obtains compelling results in challenging synthetic-to-real adaptation scenarios by transferring knowledge between monocular depth estimation and semantic segmentation tasks.

Selim Kuzucu, Alessio Tonioni, Vasile Lup et al.

Large Vision-Language Models (LVLMs) map visual inputs into dense token sequences, imposing a quadratic computational bottleneck for inference. Elastic visual-token compression addresses this by training a single model that can run at multiple visual-token budgets. However, existing approaches struggle under aggressive compression. Spatial-only compression, as in nested pooling, behaves as an imperfect low-pass filter and induces spectral aliasing that obscures fine-grained detail. Query-only compression, as in nested query resampling, replaces explicit grid-aligned tokens with non-local summaries and substantially degrades spatial grounding. To resolve this representational conflict, we introduce PARCEL (Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding), a visual tokenization architecture that dynamically partitions the labor of feature extraction. PARCEL establishes spatial pool tokens as low-frequency layout anchors and conditions elastic query tokens on these anchors through Pool-Conditioned Query Resampling. This encourages query tokens to focus on complementary visual features rather than redundant spatial mapping. Extensive evaluations across 27 benchmarks show that PARCEL improves the performance-efficiency Pareto frontier, consistently outperforming existing matryoshka baselines across visual-token budgets while preserving the "train once, deploy anywhere" paradigm.

Enis Simsar, Alessio Tonioni, Evin Pınar Örnek et al.

3D GANs have the ability to generate latent codes for entire 3D volumes rather than only 2D images. These models offer desirable features like high-quality geometry and multi-view consistency, but, unlike their 2D counterparts, complex semantic image editing tasks for 3D GANs have only been partially explored. To address this problem, we propose LatentSwap3D, a semantic edit approach based on latent space discovery that can be used with any off-the-shelf 3D or 2D GAN model and on any dataset. LatentSwap3D relies on identifying the latent code dimensions corresponding to specific attributes by feature ranking using a random forest classifier. It then performs the edit by swapping the selected dimensions of the image being edited with the ones from an automatically selected reference image. Compared to other latent space control-based edit methods, which were mainly designed for 2D GANs, our method on 3D GANs provides remarkably consistent semantic edits in a disentangled manner and outperforms others both qualitatively and quantitatively. We show results on seven 3D GANs (pi-GAN, GIRAFFE, StyleSDF, MVCGAN, EG3D, StyleNeRF, and VolumeGAN) and on five datasets (FFHQ, AFHQ, Cats, MetFaces, and CompCars).

Diego Martin Arroyo, Alessio Tonioni, Federico Tombari

Current methods for image-to-image translation produce compelling results, however, the applied transformation is difficult to control, since existing mechanisms are often limited and non-intuitive. We propose ParGAN, a generalization of the cycle-consistent GAN framework to learn image transformations with simple and intuitive controls. The proposed generator takes as input both an image and a parametrization of the transformation. We train this network to preserve the content of the input image while ensuring that the result is consistent with the given parametrization. Our approach does not require paired data and can learn transformations across several tasks and datasets. We show how, with disjoint image domains with no annotated parametrization, our framework can create smooth interpolations as well as learn multiple transformations simultaneously.

Eric Tillmann Bill, Enis Simsar, Alessio Tonioni et al.

Modern text-to-image diffusion models encode rich visual priors, but expose them only through one-way text-conditioned generation. Existing unified vision--language models derived from them recover bidirectional capability through large-scale joint pretraining or substantial retraining of the text pathway, discarding the strong image prior the text-to-image backbone already encodes. We introduce \emph{FullFlow}, a parameter-efficient recipe that upgrades a pretrained rectified-flow text-to-image model into a bidirectional vision--language generator by training only LoRA adapters and lightweight text heads. FullFlow keeps images in their native continuous flow and adds a discrete insertion process for text. Separate image and text timesteps turn inference into trajectory selection in a two-dimensional generative space, enabling text$\rightarrow$image, image$\rightarrow$text, joint sampling, and partial-text prediction with a single backbone. On Stable Diffusion 3 (SD3) under an identical trainable-parameter count and matched LoRA rank, FullFlow improves text$\rightarrow$image FID from $62.7$ to $31.6$ and image$\rightarrow$text CIDEr from $2.0$ to $99.4$ over a LoRA equivalent following the previous SOTA formulation (Dual Diffusion) at matched wall-clock training time, while reducing peak VRAM from ${\sim}84$\,GB to ${\sim}38$\,GB and raising throughput by ${\sim}8\times$ on two RTX A5000 GPUs in under 24 hours, training only ${\sim}5\%$ of the backbone parameters. The same recipe transfers to FLUX.1-dev and supports downstream VQA through partial-text generation. These results show that strong bidirectional vision--language capability can be unlocked from pretrained text-to-image flow models without full multimodal pretraining.

Chiara Plizzari, Alessio Tonioni, Yongqin Xian et al.

Understanding fine-grained temporal dynamics is crucial in egocentric videos, where continuous streams capture frequent, close-up interactions with objects. In this work, we bring to light that current egocentric video question-answering datasets often include questions that can be answered using only few frames or commonsense reasoning, without being necessarily grounded in the actual video. Our analysis shows that state-of-the-art Multi-Modal Large Language Models (MLLMs) on these benchmarks achieve remarkably high performance using just text or a single frame as input. To address these limitations, we introduce EgoTempo, a dataset specifically designed to evaluate temporal understanding in the egocentric domain. EgoTempo emphasizes tasks that require integrating information across the entire video, ensuring that models would need to rely on temporal patterns rather than static cues or pre-existing knowledge. Extensive experiments on EgoTempo show that current MLLMs still fall short in temporal reasoning on egocentric videos, and thus we hope EgoTempo will catalyze new research in the field and inspire models that better capture the complexity of temporal dynamics. Dataset and code are available at https://github.com/google-research-datasets/egotempo.git.

Jiahao Xie, Alessio Tonioni, Nathalie Rauschmayr et al.

Reinforcement learning (RL) with verifiable rewards (RLVR) has demonstrated the great potential of enhancing the reasoning abilities in multimodal large language models (MLLMs). However, the reliance on language-centric priors and expensive manual annotations prevents MLLMs' intrinsic visual understanding and scalable reward designs. In this work, we introduce SSL-R1, a generic self-supervised RL framework that derives verifiable rewards directly from images. To this end, we revisit self-supervised learning (SSL) in visual domains and reformulate widely-used SSL tasks into a set of verifiable visual puzzles for RL post-training, requiring neither human nor external model supervision. Training MLLMs on these tasks substantially improves their performance on multimodal understanding and reasoning benchmarks, highlighting the potential of leveraging vision-centric self-supervised tasks for MLLM post-training. We think this work will provide useful experience in devising effective self-supervised verifiable rewards to enable RL at scale. Project page: https://github.com/Jiahao000/SSL-R1.

Carmine Zaccagnino, Fabio Quattrini, Enis Simsar et al.

Flow matching models have recently emerged as an efficient alternative to diffusion, especially for text-guided image generation and editing, offering faster inference through continuous-time dynamics. However, existing flow-based editors predominantly support global or single-instruction edits and struggle with multi-instance scenarios, where multiple parts of a reference input must be edited independently without semantic interference. We identify this limitation as a consequence of globally conditioned velocity fields and joint attention mechanisms, which entangle concurrent edits. To address this issue, we introduce Instance-Disentangled Attention, a mechanism that partitions joint attention operations, enforcing binding between instance-specific textual instructions and spatial regions during velocity field estimation. We evaluate our approach on both natural image editing and a newly introduced benchmark of text-dense infographics with region-level editing instructions. Experimental results demonstrate that our approach promotes edit disentanglement and locality while preserving global output coherence, enabling single-pass, instance-level editing.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Jiahao Xie, Alessio Tonioni, Nathalie Rauschmayr et al.

Large vision-language models (LVLMs) have demonstrated impressive performance in various multimodal understanding and reasoning tasks. However, they still struggle with object hallucinations, i.e., the claim of nonexistent objects in the visual input. To address this challenge, we propose Region-aware Chain-of-Verification (R-CoV), a visual chain-of-verification method to alleviate object hallucinations in LVLMs in a post-hoc manner. Motivated by how humans comprehend intricate visual information -- often focusing on specific image regions or details within a given sample -- we elicit such region-level processing from LVLMs themselves and use it as a chaining cue to detect and alleviate their own object hallucinations. Specifically, our R-CoV consists of six steps: initial response generation, entity extraction, coordinate generation, region description, verification execution, and final response generation. As a simple yet effective method, R-CoV can be seamlessly integrated into various LVLMs in a training-free manner and without relying on external detection models. Extensive experiments on several widely used hallucination benchmarks across multiple LVLMs demonstrate that R-CoV can significantly alleviate object hallucinations in LVLMs. Project page: https://github.com/Jiahao000/R-CoV.

Jiahao Xie, Alessio Tonioni, Nathalie Rauschmayr et al.

Visual in-context learning (VICL), as a new paradigm in computer vision, allows the model to rapidly adapt to various tasks with only a handful of prompts and examples. While effective, the existing VICL paradigm exhibits poor generalizability under distribution shifts. In this work, we propose test-time Visual In-Context Tuning (VICT), a method that can adapt VICL models on the fly with a single test sample. Specifically, we flip the role between the task prompts and the test sample and use a cycle consistency loss to reconstruct the original task prompt output. Our key insight is that a model should be aware of a new test distribution if it can successfully recover the original task prompts. Extensive experiments on six representative vision tasks ranging from high-level visual understanding to low-level image processing, with 15 common corruptions, demonstrate that our VICT can improve the generalizability of VICL to unseen new domains. In addition, we show the potential of applying VICT for unseen tasks at test time. Code: https://github.com/Jiahao000/VICT.

Svitlana Morkva, Maximum Wilder-Smith, Michael Oechsle et al.

We present MOSAIC-GS, a novel, fully explicit, and computationally efficient approach for high-fidelity dynamic scene reconstruction from monocular videos using Gaussian Splatting. Monocular reconstruction is inherently ill-posed due to the lack of sufficient multiview constraints, making accurate recovery of object geometry and temporal coherence particularly challenging. To address this, we leverage multiple geometric cues, such as depth, optical flow, dynamic object segmentation, and point tracking. Combined with rigidity-based motion constraints, these cues allow us to estimate preliminary 3D scene dynamics during an initialization stage. Recovering scene dynamics prior to the photometric optimization reduces reliance on motion inference from visual appearance alone, which is often ambiguous in monocular settings. To enable compact representations, fast training, and real-time rendering while supporting non-rigid deformations, the scene is decomposed into static and dynamic components. Each Gaussian in the dynamic part of the scene is assigned a trajectory represented as time-dependent Poly-Fourier curve for parameter-efficient motion encoding. We demonstrate that MOSAIC-GS achieves substantially faster optimization and rendering compared to existing methods, while maintaining reconstruction quality on par with state-of-the-art approaches across standard monocular dynamic scene benchmarks.

Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar et al.

Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. "blindness" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs.

Bruno Korbar, Yongqin Xian, Alessio Tonioni et al.

In this paper we present a text-conditioned video resampler (TCR) module that uses a pre-trained and frozen visual encoder and large language model (LLM) to process long video sequences for a task. TCR localises relevant visual features from the video given a text condition and provides them to a LLM to generate a text response. Due to its lightweight design and use of cross-attention, TCR can process more than 100 frames at a time with plain attention and without optimised implementations. We make the following contributions: (i) we design a transformer-based sampling architecture that can process long videos conditioned on a task, together with a training method that enables it to bridge pre-trained visual and language models; (ii) we identify tasks that could benefit from longer video perception; and (iii) we empirically validate its efficacy on a wide variety of evaluation tasks including NextQA, EgoSchema, and the EGO4D-LTA challenge.

Mohamad Shahbazi, Liesbeth Claessens, Michael Niemeyer et al. · eth-zurich

We introduce InseRF, a novel method for generative object insertion in the NeRF reconstructions of 3D scenes. Based on a user-provided textual description and a 2D bounding box in a reference viewpoint, InseRF generates new objects in 3D scenes. Recently, methods for 3D scene editing have been profoundly transformed, owing to the use of strong priors of text-to-image diffusion models in 3D generative modeling. Existing methods are mostly effective in editing 3D scenes via style and appearance changes or removing existing objects. Generating new objects, however, remains a challenge for such methods, which we address in this study. Specifically, we propose grounding the 3D object insertion to a 2D object insertion in a reference view of the scene. The 2D edit is then lifted to 3D using a single-view object reconstruction method. The reconstructed object is then inserted into the scene, guided by the priors of monocular depth estimation methods. We evaluate our method on various 3D scenes and provide an in-depth analysis of the proposed components. Our experiments with generative insertion of objects in several 3D scenes indicate the effectiveness of our method compared to the existing methods. InseRF is capable of controllable and 3D-consistent object insertion without requiring explicit 3D information as input. Please visit our project page at https://mohamad-shahbazi.github.io/inserf.

Mauro Comi, Alessio Tonioni, Max Yang et al.

Touch and vision go hand in hand, mutually enhancing our ability to understand the world. From a research perspective, the problem of mixing touch and vision is underexplored and presents interesting challenges. To this end, we propose Tactile-Informed 3DGS, a novel approach that incorporates touch data (local depth maps) with multi-view vision data to achieve surface reconstruction and novel view synthesis. Our method optimises 3D Gaussian primitives to accurately model the object's geometry at points of contact. By creating a framework that decreases the transmittance at touch locations, we achieve a refined surface reconstruction, ensuring a uniformly smooth depth map. Touch is particularly useful when considering non-Lambertian objects (e.g. shiny or reflective surfaces) since contemporary methods tend to fail to reconstruct with fidelity specular highlights. By combining vision and tactile sensing, we achieve more accurate geometry reconstructions with fewer images than prior methods. We conduct evaluation on objects with glossy and reflective surfaces and demonstrate the effectiveness of our approach, offering significant improvements in reconstruction quality.

Enis Simsar, Alessio Tonioni, Yongqin Xian et al.

Diffusion models (DMs) have gained prominence due to their ability to generate high-quality varied images with recent advancements in text-to-image generation. The research focus is now shifting towards the controllability of DMs. A significant challenge within this domain is localized editing, where specific areas of an image are modified without affecting the rest of the content. This paper introduces LIME for localized image editing in diffusion models. LIME does not require user-specified regions of interest (RoI) or additional text input, but rather employs features from pre-trained methods and a straightforward clustering method to obtain precise editing mask. Then, by leveraging cross-attention maps, it refines these segments for finding regions to obtain localized edits. Finally, we propose a novel cross-attention regularization technique that penalizes unrelated cross-attention scores in the RoI during the denoising steps, ensuring localized edits. Our approach, without re-training, fine-tuning and additional user inputs, consistently improves the performance of existing methods in various editing benchmarks. The project page can be found at https://enisimsar.github.io/LIME/.

Vishaal Udandarao, Nikhil Parthasarathy, Muhammad Ferjad Naeem et al. · cambridge

Knowledge distillation (KD) is the de facto standard for compressing large-scale models into smaller ones. Prior works have explored ever more complex KD strategies involving different objective functions, teacher-ensembles, and weight inheritance. In this work we explore an alternative, yet simple approach -- active data curation as effective distillation for contrastive multimodal pretraining. Our simple online batch selection method, ACID, outperforms strong KD baselines across various model-, data- and compute-configurations. Further, we find such an active data curation strategy to in fact be complementary to standard KD, and can be effectively combined to train highly performant inference-efficient models. Our simple and scalable pretraining framework, ACED, achieves state-of-the-art results across 27 zero-shot classification and retrieval tasks with upto 11% less inference FLOPs. We further demonstrate that our ACED models yield strong vision-encoders for training generative multimodal models in the LiT-Decoder setting, outperforming larger vision encoders for image-captioning and visual question-answering tasks.

Vittorio Pippi, Fabio Quattrini, Silvia Cascianelli et al.

Styled Handwritten Text Generation (HTG) has recently received attention from the computer vision and document analysis communities, which have developed several solutions, either GAN- or diffusion-based, that achieved promising results. Nonetheless, these strategies fail to generalize to novel styles and have technical constraints, particularly in terms of maximum output length and training efficiency. To overcome these limitations, in this work, we propose a novel framework for text image generation, dubbed Emuru. Our approach leverages a powerful text image representation model (a variational autoencoder) combined with an autoregressive Transformer. Our approach enables the generation of styled text images conditioned on textual content and style examples, such as specific fonts or handwriting styles. We train our model solely on a diverse, synthetic dataset of English text rendered in over 100,000 typewritten and calligraphy fonts, which gives it the capability to reproduce unseen styles (both fonts and users' handwriting) in zero-shot. To the best of our knowledge, Emuru is the first autoregressive model for HTG, and the first designed specifically for generalization to novel styles. Moreover, our model generates images without background artifacts, which are easier to use for downstream applications. Extensive evaluation on both typewritten and handwritten, any-length text image generation scenarios demonstrates the effectiveness of our approach.

Carmine Zaccagnino, Fabio Quattrini, Vittorio Pippi et al.

Generating faithful and readable styled text images (especially for Styled Handwritten Text generation - HTG) is an open problem with several possible applications across graphic design, document understanding, and image editing. A lot of research effort in this task is dedicated to developing strategies that reproduce the stylistic characteristics of a given writer, with promising results in terms of style fidelity and generalization achieved by the recently proposed Autoregressive Transformer paradigm for HTG. However, this method requires additional inputs, lacks a proper stop mechanism, and might end up in repetition loops, generating visual artifacts. In this work, we rethink the autoregressive formulation by framing HTG as a multimodal prompt-conditioned generation task, and tackle the content controllability issues by introducing special textual input tokens for better alignment with the visual ones. Moreover, we devise a Classifier-Free-Guidance-based strategy for our autoregressive model. Through extensive experimental validation, we demonstrate that our approach, dubbed Eruku, compared to previous solutions requires fewer inputs, generalizes better to unseen styles, and follows more faithfully the textual prompt, improving content adherence.

Luca Zanella, Massimiliano Mancini, Yiming Wang et al.

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.

Anna Kukleva, Enis Simsar, Alessio Tonioni et al.

Most existing approaches to referring segmentation achieve strong performance only through fine-tuning or by composing multiple pre-trained models, often at the cost of additional training and architectural modifications. Meanwhile, large-scale generative diffusion models encode rich semantic information, making them attractive as general-purpose feature extractors. In this work, we introduce a new method that directly exploits features, attention scores, from diffusion transformers for downstream tasks, requiring neither architectural modifications nor additional training. To systematically evaluate these features, we extend benchmarks with vision-language grounding tasks spanning both images and videos. Our key insight is that stop words act as attention magnets: they accumulate surplus attention and can be filtered to reduce noise. Moreover, we identify global attention sinks (GAS) emerging in deeper layers and show that they can be safely suppressed or redirected onto auxiliary tokens, leading to sharper and more accurate grounding maps. We further propose an attention redistribution strategy, where appended stop words partition background activations into smaller clusters, yielding sharper and more localized heatmaps. Building on these findings, we develop RefAM, a simple training-free grounding framework that combines cross-attention maps, GAS handling, and redistribution. Across zero-shot referring image and video segmentation benchmarks, our approach consistently outperforms prior methods, establishing a new state of the art without fine-tuning or additional components.

Enis Simsar, Alessio Tonioni, Yongqin Xian et al.

We propose an unsupervised instruction-based image editing approach that removes the need for ground-truth edited images during training. Existing methods rely on supervised learning with triplets of input images, ground-truth edited images, and edit instructions. These triplets are typically generated either by existing editing methods, introducing biases, or through human annotations, which are costly and limit generalization. Our approach addresses these challenges by introducing a novel editing mechanism called Edit Reversibility Constraint (ERC), which applies forward and reverse edits in one training step and enforces alignment in image, text, and attention spaces. This allows us to bypass the need for ground-truth edited images and unlock training for the first time on datasets comprising either real image-caption pairs or image-caption-instruction triplets. We empirically show that our approach performs better across a broader range of edits with high-fidelity and precision. By eliminating the need for pre-existing datasets of triplets, reducing biases associated with current methods, and proposing ERC, our work represents a significant advancement in unblocking scaling of instruction-based image editing.

Farid Yagubbayli, Yida Wang, Alessio Tonioni et al.

Most modern deep learning-based multi-view 3D reconstruction techniques use RNNs or fusion modules to combine information from multiple images after independently encoding them. These two separate steps have loose connections and do not allow easy information sharing among views. We propose LegoFormer, a transformer model for voxel-based 3D reconstruction that uses the attention layers to share information among views during all computational stages. Moreover, instead of predicting each voxel independently, we propose to parametrize the output with a series of low-rank decomposition factors. This reformulation allows the prediction of an object as a set of independent regular structures then aggregated to obtain the final reconstruction. Experiments conducted on ShapeNet demonstrate the competitive performance of our model with respect to the state of the art while having increased interpretability thanks to the self-attention layers. We also show promising generalization results to real data.

Pierluigi Zama Ramirez, Diego Martin Arroyo, Alessio Tonioni et al.

Novel view synthesis from a single image has recently achieved remarkable results, although the requirement of some form of 3D, pose, or multi-view supervision at training time limits the deployment in real scenarios. This work aims at relaxing these assumptions enabling training of conditional generative models for novel view synthesis in a completely unsupervised manner. We first pre-train a purely generative decoder model using a 3D-aware GAN formulation while at the same time train an encoder network to invert the mapping from latent space to images. Then, we swap encoder and decoder and train the network as a conditioned GAN with a mixture of an autoencoder-like objective and self-distillation. At test time, given a view of an object, our model first embeds the image content in a latent code and regresses its pose, then generates novel views of it by keeping the code fixed and varying the pose. We test our framework on both synthetic datasets such as ShapeNet and on unconstrained collections of natural images, where no competing methods can be trained.

Mattia Segu, Alessio Tonioni, Federico Tombari

Domain generalization aims at training machine learning models to perform robustly across different and unseen domains. Several recent methods use multiple datasets to train models to extract domain-invariant features, hoping to generalize to unseen domains. Instead, first we explicitly train domain-dependant representations by using ad-hoc batch normalization layers to collect independent domain's statistics. Then, we propose to use these statistics to map domains in a shared latent space, where membership to a domain can be measured by means of a distance function. At test time, we project samples from an unknown domain into the same space and infer properties of their domain as a linear combination of the known ones. We apply the same mapping strategy at training and test time, learning both a latent representation and a powerful but lightweight ensemble model. We show a significant increase in classification accuracy over current state-of-the-art techniques on popular domain generalization benchmarks: PACS, Office-31 and Office-Caltech.

Riccardo Spezialetti, David Joseph Tan, Alessio Tonioni et al.

Estimating the 3D shape of an object from a single or multiple images has gained popularity thanks to the recent breakthroughs powered by deep learning. Most approaches regress the full object shape in a canonical pose, possibly extrapolating the occluded parts based on the learned priors. However, their viewpoint invariant technique often discards the unique structures visible from the input images. In contrast, this paper proposes to rely on viewpoint variant reconstructions by merging the visible information from the given views. Our approach is divided into three steps. Starting from the sparse views of the object, we first align them into a common coordinate system by estimating the relative pose between all the pairs. Then, inspired by the traditional voxel carving, we generate an occupancy grid of the object taken from the silhouette on the images and their relative poses. Finally, we refine the initial reconstruction to build a clean 3D model which preserves the details from each viewpoint. To validate the proposed method, we perform a comprehensive evaluation on the ShapeNet reference benchmark in terms of relative pose estimation and 3D shape reconstruction.

Matteo Poggi, Alessio Tonioni, Fabio Tosi et al.

Depth estimation from stereo images is carried out with unmatched results by convolutional neural networks trained end-to-end to regress dense disparities. Like for most tasks, this is possible if large amounts of labelled samples are available for training, possibly covering the whole data distribution encountered at deployment time. Being such an assumption systematically unmet in real applications, the capacity of adapting to any unseen setting becomes of paramount importance. Purposely, we propose a continual adaptation paradigm for deep stereo networks designed to deal with challenging and ever-changing environments. We design a lightweight and modular architecture, Modularly ADaptive Network (MADNet), and formulate Modular ADaptation algorithms (MAD, MAD++) which permit efficient optimization of independent sub-portions of the entire network. In our paradigm, the learning signals needed to continuously adapt models online can be sourced from self-supervision via right-to-left image warping or from traditional stereo algorithms. With both sources, no other data than the input images being gathered at deployment time are needed. Thus, our network architecture and adaptation algorithms realize the first real-time self-adaptive deep stereo system and pave the way for a new paradigm that can facilitate practical deployment of end-to-end architectures for dense disparity regression.

Alessio Tonioni, Matteo Poggi, Stefano Mattoccia et al.

State-of-the-art approaches to infer dense depth measurements from images rely on CNNs trained end-to-end on a vast amount of data. However, these approaches suffer a drastic drop in accuracy when dealing with environments much different in appearance and/or context from those observed at training time. This domain shift issue is usually addressed by fine-tuning on smaller sets of images from the target domain annotated with depth labels. Unfortunately, relying on such supervised labeling is seldom feasible in most practical settings. Therefore, we propose an unsupervised domain adaptation technique which does not require groundtruth labels. Our method relies only on image pairs and leverages on classical stereo algorithms to produce disparity measurements alongside with confidence estimators to assess upon their reliability. We propose to fine-tune both depth-from-stereo as well as depth-from-mono architectures by a novel confidence-guided loss function that handles the measured disparities as noisy labels weighted according to the estimated confidence. Extensive experimental results based on standard datasets and evaluation protocols prove that our technique can address effectively the domain shift issue with both stereo and monocular depth prediction architectures and outperforms other state-of-the-art unsupervised loss functions that may be alternatively deployed to pursue domain adaptation.

Daniele De Gregorio, Alessio Tonioni, Gianluca Palli et al.

In this paper, we propose Augmented Reality Semi-automatic labeling (ARS), a semi-automatic method which leverages on moving a 2D camera by means of a robot, proving precise camera tracking, and an augmented reality pen to define initial object bounding box, to create large labeled datasets with minimal human intervention. By removing the burden of generating annotated data from humans, we make the Deep Learning technique applied to computer vision, that typically requires very large datasets, truly automated and reliable. With the ARS pipeline, we created effortlessly two novel datasets, one on electromechanical components (industrial scenario) and one on fruits (daily-living scenario), and trained robustly two state-of-the-art object detectors, based on convolutional neural networks, such as YOLO and SSD. With respect to the conventional manual annotation of 1000 frames that takes us slightly more than 10 hours, the proposed approach based on ARS allows annotating 9 sequences of about 35000 frames in less than one hour, with a gain factor of about 450. Moreover, both the precision and recall of object detection is increased by about 15\% with respect to manual labeling. All our software is available as a ROS package in a public repository alongside the novel annotated datasets.

Oscar Rahnama, Tommaso Cavallari, Stuart Golodetz et al.

Obtaining highly accurate depth from stereo images in real time has many applications across computer vision and robotics, but in some contexts, upper bounds on power consumption constrain the feasible hardware to embedded platforms such as FPGAs. Whilst various stereo algorithms have been deployed on these platforms, usually cut down to better match the embedded architecture, certain key parts of the more advanced algorithms, e.g. those that rely on unpredictable access to memory or are highly iterative in nature, are difficult to deploy efficiently on FPGAs, and thus the depth quality that can be achieved is limited. In this paper, we leverage a FPGA-CPU chip to propose a novel, sophisticated, stereo approach that combines the best features of SGM and ELAS-based methods to compute highly accurate dense depth in real time. Our approach achieves an 8.7% error rate on the challenging KITTI 2015 dataset at over 50 FPS, with a power consumption of only 5W.

Pierluigi Zama Ramirez, Alessio Tonioni, Samuele Salti et al.

Recent works have proven that many relevant visual tasks are closely related one to another. Yet, this connection is seldom deployed in practice due to the lack of practical methodologies to transfer learned concepts across different training processes. In this work, we introduce a novel adaptation framework that can operate across both task and domains. Our framework learns to transfer knowledge across tasks in a fully supervised domain (e.g., synthetic data) and use this knowledge on a different domain where we have only partial supervision (e.g., real data). Our proposal is complementary to existing domain adaptation techniques and extends them to cross tasks scenarios providing additional performance gains. We prove the effectiveness of our framework across two challenging tasks (i.e., monocular depth estimation and semantic segmentation) and four different domains (Synthia, Carla, Kitti, and Cityscapes).

Alessio Tonioni, Oscar Rahnama, Thomas Joy et al.

Real world applications of stereo depth estimation require models that are robust to dynamic variations in the environment. Even though deep learning based stereo methods are successful, they often fail to generalize to unseen variations in the environment, making them less suitable for practical applications such as autonomous driving. In this work, we introduce a "learning-to-adapt" framework that enables deep stereo methods to continuously adapt to new target domains in an unsupervised manner. Specifically, our approach incorporates the adaptation procedure into the learning objective to obtain a base set of parameters that are better suited for unsupervised online adaptation. To further improve the quality of the adaptation, we learn a confidence measure that effectively masks the errors introduced during the unsupervised adaptation. We evaluate our method on synthetic and real-world stereo datasets and our experiments evidence that learning-to-adapt is, indeed beneficial for online adaptation on vastly different domains.

Alessio Tonioni, Luigi Di Stefano

Recognizing packaged grocery products based solely on appearance is still an open issue for modern computer vision systems due to peculiar challenges. Firstly, the number of different items to be recognized is huge (i.e., in the order of thousands) and rapidly changing over time. Moreover, there exist a significant domain shift between the images that should be recognized at test time, taken in stores by cheap cameras, and those available for training, usually just one or a few studio-quality images per product. We propose an end-to-end architecture comprising a GAN to address the domain shift at training time and a deep CNN trained on the samples generated by the GAN to learn an embedding of product images that enforces a hierarchy between product categories. At test time, we perform recognition by means of K-NN search against a database consisting of just one reference image per product. Experiments addressing recognition of products present in the training datasets as well as different ones unseen at training time show that our approach compares favourably to state-of-the-art methods on the grocery recognition task and generalize fairly well to similar ones.

Pierluigi Zama Ramirez, Alessio Tonioni, Luigi Di Stefano

Performance achievable by modern deep learning approaches are directly related to the amount of data used at training time. Unfortunately, the annotation process is notoriously tedious and expensive, especially for pixel-wise tasks like semantic segmentation. Recent works have proposed to rely on synthetically generated imagery to ease the training set creation. However, models trained on these kind of data usually under-perform on real images due to the well known issue of domain shift. We address this problem by learning a domain-to-domain image translation GAN to shrink the gap between real and synthetic images. Peculiarly to our method, we introduce semantic constraints into the generation process to both avoid artifacts and guide the synthesis. To prove the effectiveness of our proposal, we show how a semantic segmentation CNN trained on images from the synthetic GTA dataset adapted by our method can improve performance by more than 16% mIoU with respect to the same model trained on synthetic images.

Alessio Tonioni, Fabio Tosi, Matteo Poggi et al.

Deep convolutional neural networks trained end-to-end are the state-of-the-art methods to regress dense disparity maps from stereo pairs. These models, however, suffer from a notable decrease in accuracy when exposed to scenarios significantly different from the training set, e.g., real vs synthetic images, etc.). We argue that it is extremely unlikely to gather enough samples to achieve effective training/tuning in any target domain, thus making this setup impractical for many applications. Instead, we propose to perform unsupervised and continuous online adaptation of a deep stereo network, which allows for preserving its accuracy in any environment. However, this strategy is extremely computationally demanding and thus prevents real-time inference. We address this issue introducing a new lightweight, yet effective, deep stereo architecture, Modularly ADaptive Network (MADNet) and developing a Modular ADaptation (MAD) algorithm, which independently trains sub-portions of the network. By deploying MADNet together with MAD we introduce the first real-time self-adaptive deep stereo system enabling competitive performance on heterogeneous datasets.

Alessio Tonioni, Eugenio Serra, Luigi Di Stefano

Recognition of grocery products in store shelves poses peculiar challenges. Firstly, the task mandates the recognition of an extremely high number of different items, in the order of several thousands for medium-small shops, with many of them featuring small inter and intra class variability. Then, available product databases usually include just one or a few studio-quality images per product (referred to herein as reference images), whilst at test time recognition is performed on pictures displaying a portion of a shelf containing several products and taken in the store by cheap cameras (referred to as query images). Moreover, as the items on sale in a store as well as their appearance change frequently over time, a practical recognition system should handle seamlessly new products/packages. Inspired by recent advances in object detection and image retrieval, we propose to leverage on state of the art object detectors based on deep learning to obtain an initial productagnostic item detection. Then, we pursue product recognition through a similarity search between global descriptors computed on reference and cropped query images. To maximize performance, we learn an ad-hoc global descriptor by a CNN trained on reference images based on an image embedding loss. Our system is computationally expensive at training time but can perform recognition rapidly and accurately at test time.

Alessio Tonioni, Luigi Di Stefano

The arrangement of products in store shelves is carefully planned to maximize sales and keep customers happy. However, verifying compliance of real shelves to the ideal layout is a costly task routinely performed by the store personnel. In this paper, we propose a computer vision pipeline to recognize products on shelves and verify compliance to the planned layout. We deploy local invariant features together with a novel formulation of the product recognition problem as a sub-graph isomorphism between the items appearing in the given image and the ideal layout. This allows for auto-localizing the given image within the aisle or store and improving recognition dramatically.