Search

Vision-Language (VL) models have gained significant research focus, enabling remarkable advances in multimodal reasoning. These architectures typically comprise a vision encoder, a Large Language Model (LLM), and a projection module that aligns visual features with the LLM's representation space. Despite their success, a critical limitation persists: the vision encoding process remains decoupled from user queries, often in the form of image-related questions. Consequently, the resulting visual features may not be optimally attuned to the query-specific elements of the image. To address this, we introduce QA-ViT, a Question Aware Vision Transformer approach for multimodal reasoning, which embeds question awareness directly within the vision encoder. This integration results in dynamic visual features focusing on relevant image aspects to the posed question. QA-ViT is model-agnostic and can be incorporated efficiently into any VL architecture. Extensive experiments demonstrate the effectiveness of applying our method to various multimodal architectures, leading to consistent improvement across diverse tasks and showcasing its potential for enhancing visual and scene-text understanding.

The increasing use of transformer-based large language models brings forward the challenge of processing long sequences. In document visual question answering (DocVQA), leading methods focus on the single-page setting, while documents can span hundreds of pages. We present GRAM, a method that seamlessly extends pre-trained single-page models to the multi-page setting, without requiring computationally-heavy pretraining. To do so, we leverage a single-page encoder for local page-level understanding, and enhance it with document-level designated layers and learnable tokens, facilitating the flow of information across pages for global reasoning. To enforce our model to utilize the newly introduced document tokens, we propose a tailored bias adaptation method. For additional computational savings during decoding, we introduce an optional compression stage using our compression-transformer (C-Former),reducing the encoded sequence length, thereby allowing a tradeoff between quality and latency. Extensive experiments showcase GRAM's state-of-the-art performance on the benchmarks for multi-page DocVQA, demonstrating the effectiveness of our approach.

Reading text in real-world scenarios often requires understanding the context surrounding it, especially when dealing with poor-quality text. However, current scene text recognizers are unaware of the bigger picture as they operate on cropped text images. In this study, we harness the representative capabilities of modern vision-language models, such as CLIP, to provide scene-level information to the crop-based recognizer. We achieve this by fusing a rich representation of the entire image, obtained from the vision-language model, with the recognizer word-level features via a gated cross-attention mechanism. This component gradually shifts to the context-enhanced representation, allowing for stable fine-tuning of a pretrained recognizer. We demonstrate the effectiveness of our model-agnostic framework, CLIPTER (CLIP TExt Recognition), on leading text recognition architectures and achieve state-of-the-art results across multiple benchmarks. Furthermore, our analysis highlights improved robustness to out-of-vocabulary words and enhanced generalization in low-data regimes., Comment: Accepted for publication by ICCV 2023

Visual Question Answering (VQA) and Image Captioning (CAP), which are among the most popular vision-language tasks, have analogous scene-text versions that require reasoning from the text in the image. Despite their obvious resemblance, the two are treated independently and, as we show, yield task-specific methods that can either see or read, but not both. In this work, we conduct an in-depth analysis of this phenomenon and propose UniTNT, a Unified Text-Non-Text approach, which grants existing multimodal architectures scene-text understanding capabilities. Specifically, we treat scene-text information as an additional modality, fusing it with any pretrained encoder-decoder-based architecture via designated modules. Thorough experiments reveal that UniTNT leads to the first single model that successfully handles both task types. Moreover, we show that scene-text understanding capabilities can boost vision-language models' performance on general VQA and CAP by up to 2.69% and 0.6 CIDEr, respectively.

This paper presents final results of the Out-Of-Vocabulary 2022 (OOV) challenge. The OOV contest introduces an important aspect that is not commonly studied by Optical Character Recognition (OCR) models, namely, the recognition of unseen scene text instances at training time. The competition compiles a collection of public scene text datasets comprising of 326,385 images with 4,864,405 scene text instances, thus covering a wide range of data distributions. A new and independent validation and test set is formed with scene text instances that are out of vocabulary at training time. The competition was structured in two tasks, end-to-end and cropped scene text recognition respectively. A thorough analysis of results from baselines and different participants is presented. Interestingly, current state-of-the-art models show a significant performance gap under the newly studied setting. We conclude that the OOV dataset proposed in this challenge will be an essential area to be explored in order to develop scene text models that achieve more robust and generalized predictions., Comment: To be appeared in Text In Everything Workshop in ECCV 2022

Until recently, the number of public real-world text images was insufficient for training scene text recognizers. Therefore, most modern training methods rely on synthetic data and operate in a fully supervised manner. Nevertheless, the amount of public real-world text images has increased significantly lately, including a great deal of unlabeled data. Leveraging these resources requires semi-supervised approaches; however, the few existing methods do not account for vision-language multimodality structure and therefore suboptimal for state-of-the-art multimodal architectures. To bridge this gap, we present semi-supervised learning for multimodal text recognizers (SemiMTR) that leverages unlabeled data at each modality training phase. Notably, our method refrains from extra training stages and maintains the current three-stage multimodal training procedure. Our algorithm starts by pretraining the vision model through a single-stage training that unifies self-supervised learning with supervised training. More specifically, we extend an existing visual representation learning algorithm and propose the first contrastive-based method for scene text recognition. After pretraining the language model on a text corpus, we fine-tune the entire network via a sequential, character-level, consistency regularization between weakly and strongly augmented views of text images. In a novel setup, consistency is enforced on each modality separately. Extensive experiments validate that our method outperforms the current training schemes and achieves state-of-the-art results on multiple scene text recognition benchmarks., Comment: Code will be published upon publication

In this work, we study the problem of word-level confidence calibration for scene-text recognition (STR). Although the topic of confidence calibration has been an active research area for the last several decades, the case of structured and sequence prediction calibration has been scarcely explored. We analyze several recent STR methods and show that they are consistently overconfident. We then focus on the calibration of STR models on the word rather than the character level. In particular, we demonstrate that for attention based decoders, calibration of individual character predictions increases word-level calibration error compared to an uncalibrated model. In addition, we apply existing calibration methodologies as well as new sequence-based extensions to numerous STR models, demonstrating reduced calibration error by up to a factor of nearly 7. Finally, we show consistently improved accuracy results by applying our proposed sequence calibration method as a preprocessing step to beam-search.

We propose a framework for sequence-to-sequence contrastive learning (SeqCLR) of visual representations, which we apply to text recognition. To account for the sequence-to-sequence structure, each feature map is divided into different instances over which the contrastive loss is computed. This operation enables us to contrast in a sub-word level, where from each image we extract several positive pairs and multiple negative examples. To yield effective visual representations for text recognition, we further suggest novel augmentation heuristics, different encoder architectures and custom projection heads. Experiments on handwritten text and on scene text show that when a text decoder is trained on the learned representations, our method outperforms non-sequential contrastive methods. In addition, when the amount of supervision is reduced, SeqCLR significantly improves performance compared with supervised training, and when fine-tuned with 100% of the labels, our method achieves state-of-the-art results on standard handwritten text recognition benchmarks.

Deep generative models (e.g. GANs and VAEs) have been developed quite extensively in recent years. Lately, there has been an increased interest in the inversion of such a model, i.e. given a (possibly corrupted) signal, we wish to recover the latent vector that generated it. Building upon sparse representation theory, we define conditions that are applicable to any inversion algorithm (gradient descent, deep encoder, etc.), under which such generative models are invertible with a unique solution. Importantly, the proposed analysis is applicable to any trained model, and does not depend on Gaussian i.i.d. weights. Furthermore, we introduce two layer-wise inversion pursuit algorithms for trained generative networks of arbitrary depth, and accompany these with recovery guarantees. Finally, we validate our theoretical results numerically and show that our method outperforms gradient descent when inverting such generators, both for clean and corrupted signals.

Neural networks that are based on unfolding of an iterative solver, such as LISTA (learned iterative soft threshold algorithm), are widely used due to their accelerated performance. Nevertheless, as opposed to non-learned solvers, these networks are trained on a certain dictionary, and therefore they are inapplicable for varying model scenarios. This work introduces an adaptive learned solver, termed Ada-LISTA, which receives pairs of signals and their corresponding dictionaries as inputs, and learns a universal architecture to serve them all. We prove that this scheme is guaranteed to solve sparse coding in linear rate for varying models, including dictionary perturbations and permutations. We also provide an extensive numerical study demonstrating its practical adaptation capabilities. Finally, we deploy Ada-LISTA to natural image inpainting, where the patch-masks vary spatially, thus requiring such an adaptation.

Image interpolation, or image morphing, refers to a visual transition between two (or more) input images. For such a transition to look visually appealing, its desirable properties are (i) to be smooth; (ii) to apply the minimal required change in the image; and (iii) to seem "real", avoiding unnatural artifacts in each image in the transition. To obtain a smooth and straightforward transition, one may adopt the well-known Wasserstein Barycenter Problem (WBP). While this approach guarantees minimal changes under the Wasserstein metric, the resulting images might seem unnatural. In this work, we propose a novel approach for image morphing that possesses all three desired properties. To this end, we define a constrained variant of the WBP that enforces the intermediate images to satisfy an image prior. We describe an algorithm that solves this problem and demonstrate it using the sparse prior and generative adversarial networks.

Parsimonious representations are ubiquitous in modeling and processing information. Motivated by the recent Multi-Layer Convolutional Sparse Coding (ML-CSC) model, we herein generalize the traditional Basis Pursuit problem to a multi-layer setting, introducing similar sparse enforcing penalties at different representation layers in a symbiotic relation between synthesis and analysis sparse priors. We explore different iterative methods to solve this new problem in practice, and we propose a new Multi-Layer Iterative Soft Thresholding Algorithm (ML-ISTA), as well as a fast version (ML-FISTA). We show that these nested first order algorithms converge, in the sense that the function value of near-fixed points can get arbitrarily close to the solution of the original problem. We further show how these algorithms effectively implement particular recurrent convolutional neural networks (CNNs) that generalize feed-forward ones without introducing any parameters. We present and analyze different architectures resulting unfolding the iterations of the proposed pursuit algorithms, including a new Learned ML-ISTA, providing a principled way to construct deep recurrent CNNs. Unlike other similar constructions, these architectures unfold a global pursuit holistically for the entire network. We demonstrate the emerging constructions in a supervised learning setting, consistently improving the performance of classical CNNs while maintaining the number of parameters constant.

Despite their impressive performance, deep convolutional neural networks (CNNs) have been shown to be sensitive to small adversarial perturbations. These nuisances, which one can barely notice, are powerful enough to fool sophisticated and well performing classifiers, leading to ridiculous misclassification results. In this paper we analyze the stability of state-of-the-art deep-learning classification machines to adversarial perturbations, where we assume that the signals belong to the (possibly multi-layer) sparse representation model. We start with convolutional sparsity and then proceed to its multi-layered version, which is tightly connected to CNNs. Our analysis links between the stability of the classification to noise and the underlying structure of the signal, quantified by the sparsity of its representation under a fixed dictionary. In addition, we offer similar stability theorems for two practical pursuit algorithms, which are posed as two different deep-learning architectures - the layered Thresholding and the layered Basis Pursuit. Our analysis establishes the better robustness of the later to adversarial attacks. We corroborate these theoretical results by numerical experiments on three datasets: MNIST, CIFAR-10 and CIFAR-100.

The recently proposed multi-layer sparse model has raised insightful connections between sparse representations and convolutional neural networks (CNN). In its original conception, this model was restricted to a cascade of convolutional synthesis representations. In this paper, we start by addressing a more general model, revealing interesting ties to fully connected networks. We then show that this multi-layer construction admits a brand new interpretation in a unique symbiosis between synthesis and analysis models: while the deepest layer indeed provides a synthesis representation, the mid-layers decompositions provide an analysis counterpart. This new perspective exposes the suboptimality of previously proposed pursuit approaches, as they do not fully leverage all the information comprised in the model constraints. Armed with this understanding, we address fundamental theoretical issues, revisiting previous analysis and expanding it. Motivated by the limitations of previous algorithms, we then propose an integrated - holistic - alternative that estimates all representations in the model simultaneously, and analyze all these different schemes under stochastic noise assumptions. Inspired by the synthesis-analysis duality, we further present a Holistic Pursuit algorithm, which alternates between synthesis and analysis sparse coding steps, eventually solving for the entire model as a whole, with provable improved performance. Finally, we present numerical results that demonstrate the practical advantages of our approach.