Machine learning models have been found to learn shortcuts—unintended decision rules that are unable to generalize—undermining models' reliability. Previous works address this problem under the tenuous assumption that only a single shortcut exists in the training data. Real-world images are rife with multiple visual cues from background to texture. Key to advancing the reliability of vision systems is understanding whether existing methods can overcome multiple shortcuts or struggle in a Whac-A-Mole game, i.e., where mitigating one shortcut amplifies reliance on others. To address this shortcoming, we propose two benchmarks: 1) UrbanCars, a dataset with precisely controlled spurious cues, and 2) ImageNet-W, an evaluation set based on ImageNet for watermark, a shortcut we discovered affects nearly every modern vision model. Along with texture and background, ImageNet-W allows us to study multiple shortcuts emerging from training on natural images. We find computer vision models, including large foundation models—regardless of training set, architecture, and supervision—struggle when multiple shortcuts are present. Even methods explicitly designed to combat shortcuts struggle in a Whac-A-Mole dilemma. To tackle this challenge, we propose Last Layer Ensemble, a simple-yet-effective method to mitigate multiple shortcuts without Whac-A-Mole behavior. Our results surface multi-shortcut mitigation as an overlooked challenge critical to advancing the reliability of vision systems. The datasets and code are released: https://github.com/facebookresearch/Whac-A-Mole.

Our work focuses on addressing sample deficiency from low-density regions of data manifold in common image datasets. We leverage diffusion process based generative models to synthesize novel images from low-density regions. We observe that uniform sampling from diffusion models predominantly samples from high-density regions of the data manifold. Therefore, we modify the sampling process to guide it towards low-density regions while simulta-neously maintaining the fidelity of synthetic data. We rigorously demonstrate that our process successfully generates novel high fidelity samples from low-density regions. We further examine generated samples and show that the model does not memorize low-density data and indeed learns to generate novel samples from low-density regions.

In continual learning, the learner faces a stream of data whose distribution changes over time. Modern neural networks are known to suffer under this setting, as they quickly forget previously acquired knowledge. To address such catastrophic forgetting, many continual learning methods implement different types of experience replay, re-learning on past data stored in a small buffer known as episodic memory. In this work, we complement experience replay with a new objective that we call ``anchoring'', where the learner uses bilevel optimization to update its knowledge on the current task, while keeping intact the predictions on some anchor points of past tasks. These anchor points are learned using gradient-based optimization to maximize forgetting, which is approximated by fine-tuning the currently trained model on the episodic memory of past tasks. Experiments on several supervised learning benchmarks for continual learning demonstrate that our approach improves the standard experience replay in terms of both accuracy and forgetting metrics and for various sizes of episodic memory.

In this paper we present a deployed, scalable optical character recognition (OCR) system, which we call Rosetta , designed to process images uploaded daily at Facebook scale. Sharing of image content has become one of the primary ways to communicate information among internet users within social networks such as Facebook, and the understanding of such media, including its textual information, is of paramount importance to facilitate search and recommendation applications. We present modeling techniques for efficient detection and recognition of text in images and describe Rosetta 's system architecture. We perform extensive evaluation of presented technologies, explain useful practical approaches to build an OCR system at scale, and provide insightful intuitions as to why and how certain components work based on the lessons learnt during the development and deployment of the system.

Querying with an example image is a simple and intuitive interface to retrieve information from a visual database. Most of the research in image retrieval has focused on the task of instance-level image retrieval, where the goal is to retrieve images that contain the same object instance as the query image. In this work we move beyond instance-level retrieval and consider the task of semantic image retrieval in complex scenes, where the goal is to retrieve images that share the same semantics as the query image. We show that, despite its subjective nature, the task of semantically ranking visual scenes is consistently implemented across a pool of human annotators. We also show that a similarity based on human-annotated region-level captions is highly correlated with the human ranking and constitutes a good computable surrogate. Following this observation, we learn a visual embedding of the images where the similarity in the visual space is correlated with their semantic similarity surrogate. We further extend our model to learn a joint embedding of visual and textual cues that allows one to query the database using a text modifier in addition to the query image, adapting the results to the modifier. Finally, our model can ground the ranking decisions by showing regions that contributed the most to the similarity between pairs of images, providing a visual explanation of the similarity.

Convolutional Networks (ConvNets) have recently improved image recognition performance thanks to end-to-end learning of deep feed-forward models from raw pixels. Deep learning is a marked departure from the previous state of the art, the Fisher Vector (FV), which relied on gradient-based encoding of local hand-crafted features. In this paper, we discuss a novel connection between these two approaches. First, we show that one can derive gradient representations from ConvNets in a similar fashion to the FV. Second, we show that this gradient representation actually corresponds to a structured matrix that allows for efficient similarity computation. We experimentally study the benefits of transferring this representation over the outputs of ConvNet layers, and find consistent improvements on the Pascal VOC 2007 and 2012 datasets.

This paper addresses the problem of learning word image representations: given the cropped image of a word, we are interested in finding a descriptive, robust, and compact fixed-length representation. Machine learning techniques can then be supplied with these representations to produce models useful for word retrieval or recognition tasks. Although many works have focused on the machine learning aspect once a global representation has been produced, little work has been devoted to the construction of those base image representations: most works use standard coding and aggregation techniques directly on top of standard computer vision features such as SIFT or HOG. We propose to learn local mid-level features suitable for building word image representations. These features are learnt by leveraging character bounding box annotations on a small set of training images. However, contrary to other approaches that use character bounding box information, our approach does not rely on detecting the individual characters explicitly at testing time. Our local midlevel features can then be aggregated to produce a global word image signature. When pairing these features with the recent word attributes framework of [4], we obtain results comparable with or better than the state-of-the-art on matching and recognition tasks using global descriptors of only 96 dimensions.

The goal of this work is to bring semantics into the tasks of text recognition and retrieval in natural images. Although text recognition and retrieval have received a lot of attention in recent years, previous works have focused on recognizing or retrieving exactly the same word used as a query, without taking the semantics into consideration. In this paper, we ask the following question: can we predict semantic concepts directly from a word image, without explicitly trying to transcribe the word image or its characters at any point? For this goal we propose a convolutional neural network (CNN) with a weighted ranking loss objective that ensures that the concepts relevant to the query image are ranked ahead of those that are not relevant. This can also be interpreted as learning a Euclidean space where word images and concepts are jointly embedded. This model is learned in an end-to-end manner, from image pixels to semantic concepts, using a dataset of synthetically generated word images and concepts mined from a lexical database (WordNet). Our results show that, despite the complexity of the task, word images and concepts can indeed be associated with a high degree of accuracy.

This paper addresses the problems of word spotting and word recognition on images. In word spotting, the goal is to find all instances of a query word in a dataset of images. In recognition, the goal is to recognize the content of the word image, usually aided by a dictionary or lexicon. We describe an approach in which both word images and text strings are embedded in a common vectorial subspace. This is achieved by a combination of label embedding and attributes learning, and a common subspace regression. In this subspace, images and strings that represent the same word are close together, allowing one to cast recognition and retrieval tasks as a nearest neighbor problem. Contrary to most other existing methods, our representation has a fixed length, is low dimensional, and is very fast to compute and, especially, to compare. We test our approach on four public datasets of both handwritten documents and natural images showing results comparable or better than the state-of-the-art on spotting and recognition tasks.

In large-scale query-by-example retrieval, embedding image signatures in a binary space offers two benefits: data compression and search efficiency. While most embedding algorithms binarize both query and database signatures, it has been noted that this is not strictly a requirement. Indeed, asymmetric schemes that binarize the database signatures but not the query still enjoy the same two benefits but may provide superior accuracy. In this work, we propose two general asymmetric distances that are applicable to a wide variety of embedding techniques including locality sensitive hashing (LSH), locality sensitive binary codes (LSBC), spectral hashing (SH), PCA embedding (PCAE), PCAE with random rotations (PCAE-RR), and PCAE with iterative quantization (PCAE-ITQ). We experiment on four public benchmarks containing up to 1M images and show that the proposed asymmetric distances consistently lead to large improvements over the symmetric Hamming distance for all binary embedding techniques.

We propose an approach to multi-writer word spotting, where the goal is to find a query word in a dataset comprised of document images. We propose an attributes-based approach that leads to a low-dimensional, fixed-length representation of the word images that is fast to compute and, especially, fast to compare. This approach naturally leads to an unified representation of word images and strings, which seamlessly allows one to indistinctly perform query-by-example, where the query is an image, and query-by-string, where the query is a string. We also propose a calibration scheme to correct the attributes scores based on Canonical Correlation Analysis that greatly improves the results on a challenging dataset. We test our approach on two public datasets showing state-of-the-art results.

In this paper we present a method for the segmentation of continuous page streams into multipage documents and the simultaneous classification of the resulting documents. We first present an approach to combine the multiple pages of a document into a single feature vector that represents the whole document. Despite its simplicity and low computational cost, the proposed representation yields results comparable to more complex methods in multipage document classification tasks. We then exploit this representation in the context of page stream segmentation. The most plausible segmentation of a page stream into a sequence of multipage documents is obtained by optimizing a statistical model that represents the probability of each segmented multipage document belonging to a particular class. Experimental results are reported on a large sample of real administrative multipage documents.

This paper addresses the problem of learning similarity-preserving binary codes for efficient similarity search in large-scale image collections. We formulate this problem in terms of finding a rotation of zero-centered data so as to minimize the quantization error of mapping this data to the vertices of a zero-centered binary hypercube, and propose a simple and efficient alternating minimization algorithm to accomplish this task. This algorithm, dubbed iterative quantization (ITQ), has connections to multiclass spectral clustering and to the orthogonal Procrustes problem, and it can be used both with unsupervised data embeddings such as PCA and supervised embeddings such as canonical correlation analysis (CCA). The resulting binary codes significantly outperform several other state-of-the-art methods. We also show that further performance improvements can result from transforming the data with a nonlinear kernel mapping prior to PCA or CCA. Finally, we demonstrate an application of ITQ to learning binary attributes or "classemes" on the ImageNet data set.

In this paper we propose an unsupervised segmentation-free method for word spotting in document images. Documents are represented with a grid of HOG descriptors, and a sliding window approach is used to locate the document regions that are most similar to the query. We use the Exemplar SVM framework to produce a better representation of the query in an unsupervised way. Finally, the document descriptors are precomputed and compressed with Product Quantization. This offers two advantages: first, a large number of documents can be kept in RAM memory at the same time. Second, the sliding window becomes significantly faster since distances between quantized HOG descriptors can be precomputed. Our results significantly outperform other segmentation-free methods in the literature, both in accuracy and in speed and memory usage.

In this article, we focus on the problem of large-scale instance-level image retrieval. For efficiency reasons, it is common to represent an image by a fixed-length descriptor which is subsequently encoded into a small number of bits. We note that most encoding techniques include an unsupervised dimensionality reduction step. Our goal in this work is to learn a better subspace in a supervised manner. We especially raise the following question: "can category-level labels be used to learn such a subspace?" To answer this question, we experiment with four learning techniques: the first one is based on a metric learning framework, the second one on attribute representations, the third one on Canonical Correlation Analysis (CCA) and the fourth one on Joint Subspace and Classifier Learning (JSCL). While the first three approaches have been applied in the past to the image retrieval problem, we believe we are the first to show the usefulness of JSCL in this context. In our experiments, we use ImageNet as a source of category-level labels and report retrieval results on two standard dataseis: INRIA Holidays and the University of Kentucky benchmark. Our experimental study shows that metric learning and attributes do not lead to any significant improvement in retrieval accuracy, as opposed to CCA and JSCL. As an example, we report on Holidays an increase in accuracy from 39.3% to 48.6% with 32-dimensional representations. Overall JSCL is shown to yield the best results.

The combination of multiple features or views when representing documents or other kinds of objects usually leads to improved results in classification (and retrieval) tasks. Most systems assume that those views will be available both at training and test time. However, some views may be too `expensive' to be available at test time. In this paper, we consider the use of Canonical Correlation Analysis to leverage `expensive' views that are available only at training time. Experimental results show that this information may significantly improve the results in a classification task.

In large-scale query-by-example retrieval, embedding image signatures in a binary space offers two benefits: data compression and search efficiency. While most embedding algorithms binarize both query and database signatures, it has been noted that this is not strictly a requirement. Indeed, asymmetric schemes which binarize the database signatures but not the query still enjoy the same two benefits but may provide superior accuracy. In this work, we propose two general asymmetric distances which are applicable to a wide variety of embedding techniques including Locality Sensitive Hashing (LSH), Locality Sensitive Binary Codes (LSBC), Spectral Hashing (SH) and Semi-Supervised Hashing (SSH). We experiment on four public benchmarks containing up to 1M images and show that the proposed asymmetric distances consistently lead to large improvements over the symmetric Hamming distance for all binary embedding techniques. We also propose a novel simple binary embedding technique - PCA Embedding (PCAE) - which is shown to yield competitive results with respect to more complex algorithms such as SH and SSH.

In the last years, there has been a growing interest in the analysis of handwritten music scores. In this sense, our goal has been to foster the interest in the analysis of handwritten music scores by the proposal of two different competitions: Staff removal and Writer Identification. Both competitions have been tested on the CVC-MUSCIMA database: a ground-truth of handwritten music score images. This paper describes the competition details, including the dataset and ground-truth, the evaluation metrics, and a short description of the participants, their methods, and the obtained results.

Determining the authorship of a document, namely writer identification, can be an important source of information for document categorization. Contrary to text documents, the identification of the writer of graphical documents is still a challenge. In this paper we present a robust approach for writer identification in a particular kind of graphical documents, old music scores. This approach adapts the bag of visual terms method for coping with graphic documents. The identification is performed only using the graphical music notation. For this purpose, we generate a graphic vocabulary without recognizing any music symbols, and consequently, avoiding the difficulties in the recognition of hand-drawn symbols in old and degraded documents. The proposed method has been tested on a database of old music scores from the 17th to 19th centuries, achieving very high identification rates.

In this paper we tackle the problem of document image retrieval by combining a similarity measure between documents and the probability that a given document belongs to a certain class. The membership probability to a specific class is computed using Support Vector Machines in conjunction with similarity measure based kernel applied to structural document representations. In the presented experiments, we use different document representations, both visual and structural, and we apply them to a database of historical documents. We show how our method based on similarity kernels outperforms the usual distance-based retrieval.

We consider the problem of classifying documents containing multiple unordered pages. For this purpose, we propose a novel bag-of-pages document representation. To represent a document, one assigns every page to a prototype in a codebook of pages. This leads to a histogram representation which can then be fed to any discriminative classifier. We also consider several refinements over this initial approach. We show on two challenging datasets that the proposed approach significantly outperforms a baseline system.

Document classification usually requires of structural features such as the physical layout to obtain good accuracy rates on complex documents. This paper introduces a descriptor of the layout and a distance measure based on the cyclic dynamic time warping which can be computed in O(n2). This descriptor is translation invariant and can be easily modified to be scale and rotation invariant. Experiments with this descriptor and its rotation invariant modification are performed on the Girona archives database and compared against another common layout distance, the minimum weight edge cover. The experiments show that these methods outperform the MWEC both in accuracy and speed, particularly on rotated documents.

We present a complete assisted transcription system for ancient documents: State. The system consists of two applications: a pen-based, interactive application to assist humans in transcribing ancient documents and a recognition engine which offers automatic transcriptions via a web service. The interaction model and the recognition algorithm employed in the current version of State are presented. Some preliminary experiments show the productivity gains obtained with the system when transcribing a document and the error rate of the current recognition engine.