The progress of composed image retrieval (CIR), a popular research direction in image retrieval, where a combined visual and textual query is used, is held back by the absence of high-quality training and evaluation data. We introduce a new evaluation dataset, i-CIR, which, unlike existing datasets, focuses on an instance-level class definition. The goal is to retrieve images that contain the same particular object as the visual query, presented under a variety of modifications defined by textual queries. Its design and curation process keep the dataset compact to facilitate future research, while maintaining its challenge—comparable to retrieval among more than 40M random distractors—through a semi-automated selection of hard negatives. To overcome the challenge of obtaining clean, diverse, and suitable training data, we leverage pre-trained vision-and-language models (VLMs) in a training-free approach called BASIC. The method separately estimates query-image-to-image and query-text-to-image similarities, performing late fusion to upweight images that satisfy both queries, while down-weighting those that exhibit high similarity with only one of the two. Each individual similarity is further improved by a set of components that are simple and intuitive. BASIC sets a new state of the art on i-CIR but also on existing CIR datasets that follow a semantic-level class definition. Project page: https://vrg.fel.cvut.cz/icir/.

This work addresses composed image retrieval in the context of domain conversion, where the content of a query image is retrieved in the domain specified by the query text. We show that a strong vision-language model provides sufficient descriptive power without additional training. The query image is mapped to the text input space using textual inversion. Unlike common practice that invert in the continuous space of text tokens, we use the discrete word space via a nearest-neighbor search in a text vocabulary. With this inversion, the image is softly mapped across the vocabulary and is made more robust using retrieval-based augmentation. Database images are retrieved by a weighted ensemble of text queries combining mapped words with the domain text. Our method outperforms prior art by a large margin on standard and newly introduced benchmarks. Code: https://github.com/NikosEfth/freedom

Few-shot learning investigates how to solve novel tasks given limited labeled data. Exploiting unlabeled data along with the limited labeled has shown substantial improvement in performance. In this work we propose a novel algorithm that exploits unlabeled data in order to improve the performance of few-shot learning. We focus on transductive few-shot inference, where the entire test set is available at inference time, and semi-supervised few-shot learning where unlabeled data are available and can be exploited. Our algorithm starts by leveraging the manifold structure of the labeled and unlabeled data in order to assign accurate pseudo-labels to the unlabeled data. Iteratively, it selects the most confident pseudo-labels and treats them as labeled improving the quality of pseudo-labels at every iteration. Our method surpasses or matches the state of the art results on four benchmark datasets, namely mini ImageNet, tiered ImageNet, CUB and CIFAR-FS, while being robust over feature pre-processing and the quantity of available unlabeled data. Furthermore, we investigate the setting where the unlabeled data contains data from distractor classes and propose ideas to adapt our algorithm achieving new state of the art performance in the process. Specifically, we utilize the unnormalized manifold class similarities obtained from label propagation for pseudo-label cleaning and exploit the uneven pseudo-label distribution between classes to remove noisy data. The publicly available source code can be found at https://github.com/MichalisLazarou/iLPC . • Unlabeled examples can improve few-shot learning performance significantly. • A novel algorithm for transductive and semi-supervised few-shot learning is proposed. • Accurate predictions for unlabeled examples are obtained by label propagation. • Label cleaning is used to select the most confident pseudo-labeled examples. • State of the art performance achieved in multiple few-shot learning settings.

How important is it for training and evaluation sets to not have class overlap in image retrieval? We revisit Google Landmarks v2 clean [56], the most popular training set, by identifying and removing class overlap with Revisited Oxford and Paris [34], the most popular evaluation set. By comparing the original and the new RGLDv2-clean on a benchmark of reproduced state-of-the-art methods, our findings are striking. Not only is there a dramatic drop in performance, but it is inconsistent across methods, changing the ranking. What does it take to focus on objects or interest and ignore background clutter when indexing? Do we need to train an object detector and the representation separately? Do we need location supervision? We introduce Single-stage Detect-to-Retrieve (CiDeR), an end-to-end, single-stage pipeline to detect objects of interest and extract a global image representation. We outperform previous state-of-the-art on both existing training sets and the new RGLDv2-clean. Our dataset is available at https://github.com/dealicious-inc/RGLDv2-clean.

Self-supervised learning has unlocked the potential of scaling up pretraining to billions of images, since annotation is unnecessary. But are we making the best use of data? How more economical can we be? In this work, we attempt to answer this question by making two contributions. First, we investigate first-person videos and introduce a ``Walking Tours'' dataset. These videos are high-resolution, hours-long, captured in a single uninterrupted take, depicting a large number of objects and actions with natural scene transitions. They are unlabeled and uncurated, thus realistic for self-supervision and comparable with human learning. Second, we introduce a novel self-supervised image pretraining method tailored for learning from continuous videos. Existing methods typically adapt image-based pretraining approaches to incorporate more frames. Instead, we advocate a ``tracking to learn to recognize'' approach. Our method called DoRA, leads to attention maps that **D**isc**O**ver and t**RA**ck objects over time in an end-to-end manner, using transformer cross-attention. We derive multiple views from the tracks and use them in a classical self-supervised distillation loss. Using our novel approach, a single Walking Tours video remarkably becomes a strong competitor to ImageNet for several image and video downstream tasks.

This work introduces composed image retrieval to remote sensing. It allows to query a large image archive by image examples alternated by a textual description, enriching the descriptive power over unimodal queries, either visual or textual. Various attributes can be modified by the textual part, such as shape, color, or context. A novel method fusing image-to-image and text-to-image similarity is introduced. We demonstrate that a vision-language model possesses sufficient descriptive power and no further learning step or training data are necessary. We present a new evaluation benchmark focused on color, context, density, existence, quantity, and shape modifications. Our work not only sets the state-of-the-art for this task, but also serves as a foundational step in addressing a gap in the field of remote sensing image retrieval. Code at: https://github.com/billpsomas/rscir.

Transductive few-shot learning algorithms have showed substantially superior performance over their inductive counterparts by leveraging the unlabeled queries at inference. However, the vast majority of transductive methods are evaluated on perfectly class-balanced benchmarks. It has been shown that they undergo remarkable drop in performance under a more realistic, imbalanced setting.To this end, we propose a novel algorithm to address imbalanced transductive few-shot learning, named Adaptive Manifold. Our algorithm exploits the underlying manifold of the labeled examples and unlabeled queries by using manifold similarity to predict the class probability distribution of every query. It is parameterized by one centroid per class and a set of manifold parameters that determine the manifold. All parameters are optimized by minimizing a loss function that can be tuned towards class-balanced or imbalanced distributions. The manifold similarity shows substantial improvement over Euclidean distance, especially in the 1-shot setting.Our algorithm outperforms all other state of the art methods in three benchmark datasets, namely miniImageNet, tieredImageNet and CUB, and two different backbones, namely ResNet-18 and WideResNet-28-10. In certain cases, our algorithm outperforms the previous state of the art by as much as 4.2%. The publicly available source code can be found in https://github.com/MichalisLazarou/AM

Methods based on class activation maps (CAM) provide a simple mechanism to interpret predictions of convolutional neural networks by using linear combinations of feature maps as saliency maps. By contrast, masking-based methods optimize a saliency map directly in the image space or learn it by training another network on additional data. In this work we introduce Opti-CAM, combining ideas from CAM-based and masking-based approaches. Our saliency map is a linear combination of feature maps, where weights are optimized per image such that the logit of the masked image for a given class is maximized. We also fix a fundamental flaw in two of the most common evaluation metrics of attribution methods. On several datasets, Opti-CAM largely outperforms other CAM-based approaches according to the most relevant classification metrics. We provide empirical evidence supporting that localization and classifier interpretability are not necessarily aligned.

Multimodal sentiment analysis (MSA) leverages heterogeneous data sources to interpret the complex nature of human sentiments. Despite significant progress in multimodal architecture design, the field lacks comprehensive regularization methods. This paper introduces$\mathcal {P}$owMix, a versatile embedding space regularizer that builds upon the strengths of unimodal mixing-based regularization approaches and introduces novel algorithmic components that are specifically tailored to multimodal tasks.$\mathcal {P}$owMix is integrated before the fusion stage of multimodal architectures and facilitates intra-modal mixing, such as mixing text with text, to act as a regularizer.$\mathcal {P}$owMix consists of five components: 1) a varying number of generated mixed examples, 2) mixing factor reweighting, 3) anisotropic mixing, 4) dynamic mixing, and 5) cross-modal label mixing. Extensive experimentation across benchmark MSA datasets and a broad spectrum of diverse architectural designs demonstrate the efficacy of$\mathcal {P}$owMix, as evidenced by consistent performance improvements over baselines and existing mixing methods. An in-depth ablation study highlights the critical contribution of each$\mathcal {P}$owMix component and how they synergistically enhance performance. Furthermore, algorithmic analysis demonstrates how$\mathcal {P}$owMix behaves in different scenarios, particularly comparing early versus late fusion architectures. Notably,$\mathcal {P}$owMix enhances overall performance without sacrificing model robustness or magnifying text dominance. It also retains its strong performance in situations of limited data. Our findings position$\mathcal {P}$owMix as a promising versatile regularization strategy for MSA.

Impressive progress in generative models and implicit representations gave rise to methods that can generate 3D shapes of high quality. However, being able to locally con-trol and edit shapes is another essential property that can unlock several content creation applications. Local control can be achieved with part-aware models, but existing meth-ods require 3D supervision and cannot produce textures. In this work, we devise PartNeRF, a novel part-aware gener-ative model for editable 3D shape synthesis that does not require any explicit 3D supervision. Our model generates objects as a set of locally defined NeRFs, augmented with an affine transformation. This enables several editing op-erations such as applying transformations on parts, mixing parts from different objects etc. To ensure distinct, manip-ulable parts we enforce a hard assignment of rays to parts that makes sure that the color of each ray is only determined by a single NeRF. As a result, altering one part does not af-fect the appearance of the others. Evaluations on various ShapeNet categories demonstrate the ability of our model to generate editable 3D objects of improved fidelity, compared to previous part-based generative approaches that require 3D supervision or models relying on NeRFs.

Convolutional networks and vision transformers have different forms of pairwise interactions, pooling across layers and pooling at the end of the network. Does the latter really need to be different? As a by-product of pooling, vision transformers provide spatial attention for free, but this is most often of low quality unless self-supervised, which is not well studied. Is supervision really the problem?In this work, we develop a generic pooling framework and then we formulate a number of existing methods as instantiations. By discussing the properties of each group of methods, we derive SimPool, a simple attention-based pooling mechanism as a replacement of the default one for both convolutional and transformer encoders. We find that, whether supervised or self-supervised, this improves performance on pre-training and downstream tasks and provides attention maps delineating object boundaries in all cases. One could thus call SimPool universal. To our knowledge, we are the first to obtain attention maps in supervised transformers of at least as good quality as self-supervised, without explicit losses or modifying the architecture. Code at: https://github.com/billpsomas/simpool.

Few-shot learning addresses the issue of classifying images using limited labeled data. Exploiting unlabeled data through the use of transductive inference methods such as label propagation has been shown to improve the performance of few-shot learning significantly. Label propagation infers pseudo-labels for unlabeled data by utilizing a constructed graph that exploits the underlying manifold structure of the data. However, a limitation of the existing label propagation approaches is that the positions of all data points are fixed and might be sub-optimal so that the algorithm is not as effective as possible. In this work, we propose a novel algorithm that adapts the feature embeddings of the labeled data by minimizing a differentiable loss function optimizing their positions in the manifold in the process. Our novel algorithm, Adaptive Anchor Label Propagation, outperforms the standard label propagation algorithm by as much as 7% and 2% in the 1-shot and 5-shot settings respectively. We provide experimental results highlighting the merits of our algorithm on four widely used few-shot benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS and two commonly used backbones, ResNet12 and WideResNet-28-10. The source code can be found at https://github.com/MichalisLazarou/A2LP.

Mixup refers to interpolation-based data augmentation, originally motivated as a way to go beyond empirical risk minimization (ERM). Its extensions mostly focus on the definition of interpolation and the space (input or feature) where it takes place, while the augmentation process itself is less studied. In most methods, the number of generated examples is limited to the mini-batch size and the number of examples being interpolated is limited to two (pairs), in the input space. We make progress in this direction by introducing MultiMix, which generates an arbitrarily large number of interpolated examples beyond the mini-batch size and interpolates the entire mini-batch in the embedding space. Effectively, we sample on the entire convex hull of the mini-batch rather than along linear segments between pairs of examples. On sequence data, we further extend to Dense MultiMix. We densely interpolate features and target labels at each spatial location and also apply the loss densely. To mitigate the lack of dense labels, we inherit labels from examples and weight interpolation factors by attention as a measure of confidence. Overall, we increase the number of loss terms per mini-batch by orders of magnitude at little additional cost. This is only possible because of interpolating in the embedding space. We empirically show that our solutions yield significant improvement over state-of-the-art mixup methods on four different benchmarks, despite interpolation being only linear. By analyzing the embedding space, we show that the classes are more tightly clustered and uniformly spread over the embedding space, thereby explaining the improved behavior.

Vision transformers have achieved remarkable progress in vision tasks such as image classification and detection. However, in instance-level image retrieval, transformers have not yet shown good performance compared to convolutional networks. We propose a number of improvements that make transformers outperform the state of the art for the first time. (1) We show that a hybrid architecture is more effective than plain transformers, by a large margin. (2) We introduce two branches collecting global (classification token) and local (patch tokens) information, from which we form a global image representation. (3) In each branch, we collect multilayer features from the transformer encoder, corresponding to skip connections across distant layers. (4) We enhance locality of interactions at the deeper layers of the encoder, which is the relative weakness of vision transformers. We train our model on all commonly used training sets and, for the first time, we make fair comparisons separately per training set. In all cases, we outperform previous models based on global representation. Public code is available at https://github.com/dealicious-inc/DToP.

Mixup is a powerful data augmentation method that in-terpolates between two or more examples in the input or feature space and between the corresponding target labels. However, how to best interpolate images is not well defined. Recent mixup methods overlay or cut-and-paste two or more objects into one image, which needs care in selecting regions. Mixup has also been connected to autoencoders, because often autoencoders generate an image that continuously deforms into another. However, such images are typically of low quality. In this work, we revisit mixup from the deformation perspective and introduce AligtiMixup, where we geometrically align two images in the feature space. The correspondences allow us to interpolate between two sets of features, while keeping the locations of one set. Interestingly, this retains mostly the geometry or pose of one image and the appearance or texture of the other. We also show that an autoencoder can still improve representation learning under mixup, without the classifier ever seeing decoded images. AlignMixup outperforms state-of-the-art mixup methods on five different benchmarks. Code available at https://github.com/shashankvkt/AlignMixup_CVPR22.git

Few-shot learning addresses the challenge of learning how to address novel tasks given not just limited supervision but limited data as well. An attractive solution is synthetic data generation. However, most such methods are overly sophisticated, focusing on high-quality, realistic data in the input space. It is unclear whether adapting them to the few-shot regime and using them for the downstream task of classification is the right approach. Previous works on synthetic data generation for few-shot classification focus on exploiting complex models, e.g. a Wasserstein GAN with multiple regularizers or a network that transfers latent diversities from known to novel classes.We follow a different approach and investigate how a simple and straightforward synthetic data generation method can be used effectively. We make two contributions, namely we show that: (1) using a simple loss function is more than enough for training a feature generator in the few-shot setting; and (2) learning to generate tensor features instead of vector features is superior. Extensive experiments on miniImagenet, CUB and CIFAR-FS datasets show that our method sets a new state of the art, outperforming more sophisticated few-shot data augmentation methods. The source code can be found at https://github.com/MichalisLazarou/TFH_fewshot.

We address representation learning for large-scale instance-level image retrieval. Apart from backbone, training pipelines and loss functions, popular approaches have focused on different spatial pooling and attention mechanisms, which are at the core of learning a powerful global image representation. There are different forms of attention according to the interaction of elements of the feature tensor (local and global) and the dimensions where it is applied (spatial and channel). Unfortunately, each study addresses only one or two forms of attention and applies it to different problems like classification, detection or retrieval.We present global-local attention module (GLAM), which is attached at the end of a backbone network and incorporates all four forms of attention: local and global, spatial and channel. We obtain a new feature tensor and, by spatial pooling, we learn a powerful embedding for image retrieval. Focusing on global descriptors, we provide empirical evidence of the interaction of all forms of attention and improve the state of the art on standard benchmarks.

Knowledge transfer from large teacher models to smaller student models has recently been studied for metric learning, focusing on fine-grained classification. In this work, focusing on instance-level image retrieval, we study an asymmetric testing task, where the database is represented by the teacher and queries by the student. Inspired by this task, we introduce asymmetric metric learning, a novel paradigm of using asymmetric representations at training. This acts as a simple combination of knowledge transfer with the original metric learning task.We systematically evaluate different teacher and student models, metric learning and knowledge transfer loss functions on the new asymmetric testing as well as the standard symmetric testing task, where database and queries are represented by the same model. We find that plain regression is surprisingly effective compared to more complex knowledge transfer mechanisms, working best in asymmetric testing. Interestingly, our asymmetric metric learning approach works best in symmetric testing, allowing the student to even out-perform the teacher.Our implementation is publicly available,1including trained student models for all loss functions and all pairs of teacher/student models.2

High-level understanding of stories in video such as movies and TV shows from raw data is extremely challenging. Modern video question answering (VideoQA) systems often use additional human-made sources like plot synopses, scripts, video descriptions or knowledge bases. In this work, we present a new approach to understand the whole story without such external sources. The secret lies in the dialog: unlike any prior work, we treat dialog as a noisy source to be converted into text description via dialog summarization, much like recent methods treat video. The input of each modality is encoded by transformers independently, and a simple fusion method combines all modalities, using soft temporal attention for localization over long inputs. Our model outperforms the state of the art on the KnowIT VQA dataset by a large margin, without using question-specific human annotation or human-made plot summaries. It even outperforms human evaluators who have never watched any whole episode before. Code is available at https://engindeniz.github.io/dialogsummary-videoqa

Few-shot learning amounts to learning representations and acquiring knowledge such that novel tasks may be solved with both supervision and data being limited. Improved performance is possible by transductive inference, where the entire test set is available concurrently, and semi-supervised learning, where more unlabeled data is available.Focusing on these two settings, we introduce a new algorithm that leverages the manifold structure of the labeled and unlabeled data distribution to predict pseudo-labels, while balancing over classes and using the loss value distribution of a limited-capacity classifier to select the cleanest labels, iteratively improving the quality of pseudo-labels. Our solution surpasses or matches the state of the art results on four benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS, while being robust over feature space pre-processing and the quantity of available data. The publicly available source code can be found in https://github.com/MichalisLazarou/iLPC

Adversarial examples of deep neural networks are receiving ever increasing attention because they help in understanding and reducing the sensitivity to their input. This is natural given the increasing applications of deep neural networks in our everyday lives. When white-box attacks are almost always successful, it is typically only the distortion of the perturbations that matters in their evaluation. In this work, we argue that speed is important as well, especially when considering that fast attacks are required by adversarial training. Given more time, iterative methods can always find better solutions. We investigate this speed-distortion trade-off in some depth and introduce a new attack called boundary projection (BP) that improves upon existing methods by a large margin. Our key idea is that the classification boundary is a manifold in the image space: we therefore quickly reach the boundary and then optimize distortion on this manifold.

Few-shot learning is often motivated by the ability of humans to learn new tasks from few examples. However, standard few-shot classification benchmarks assume that the representation is learned on a limited amount of base class data, ignoring the amount of prior knowledge that a human may have accumulated before learning new tasks. At the same time, even if a powerful representation is available, it may happen in some domain that base class data are limited or non-existent. This motivates us to study a problem where the representation is obtained from a classifier pre-trained on a large-scale dataset of a different domain, assuming no access to its training process, while the base class data are limited to few examples per class and their role is to adapt the representation to the domain at hand rather than learn from scratch. We adapt the representation in two stages, namely on the few base class data if available and on the even fewer data of new tasks. In doing so, we obtain from the pre-trained classifier a spatial attention map that allows focusing on objects and suppressing background clutter. This is important in the new problem, because when base class data are few, the network cannot learn where to focus implicitly. We also show that a pre-trained network may be easily adapted to novel classes, without meta-learning.

The challenge in few-shot learning is that available data is not enough to capture the underlying distribution. To mitigate this, two emerging directions are (a) using local image representations, essentially multiplying the amount of data by a constant factor, and (b) using more unlabeled data, for instance by transductive inference, jointly on a number of queries. In this work, we bring these two ideas together, introducing local propagation. We treat local image features as independent examples, we build a graph on them and we use it to propagate both the features themselves and the labels, known and unknown. Interestingly, since there is a number of features per image, even a single query gives rise to transductive inference. As a result, we provide a universally safe choice for few-shot inference under both non-transductive and transductive settings, improving accuracy over corresponding methods. This is in contrast to existing solutions, where one needs to choose the method depending on the quantity of available data.

Active learning typically focuses on training a model on few labeled examples alone, while unlabeled ones are only used for acquisition. In this work we depart from this setting by using both labeled and unlabeled data during model training across active learning cycles. We do so by using unsupervised feature learning at the beginning of the active learning pipeline and semi-supervised learning at every active learning cycle, on all available data. The former has not been investigated before in active learning, while the study of latter in the context of deep learning is scarce and recent findings are not conclusive with respect to its benefit. Our idea is orthogonal to acquisition strategies by using more data, much like ensemble methods use more models. By systematically evaluating on a number of popular acquisition strategies and datasets, we find that the use of unlabeled data during model training brings a spectacular accuracy improvement in image classification, compared to the differences between acquisition strategies. We thus explore smaller label budgets, even one label per class.

Abstract This paper investigates the visual quality of the adversarial examples. Recent papers propose to smooth the perturbations to get rid of high frequency artifacts. In this work, smoothing has a different meaning as it perceptually shapes the perturbation according to the visual content of the image to be attacked. The perturbation becomes locally smooth on the flat areas of the input image, but it may be noisy on its textured areas and sharp across its edges.This operation relies on Laplacian smoothing, well-known in graph signal processing, which we integrate in the attack pipeline. We benchmark several attacks with and without smoothing under a white box scenario and evaluate their transferability. Despite the additional constraint of smoothness, our attack has the same probability of success at lower distortion.

Semi-supervised learning is becoming increasingly important because it can combine data carefully labeled by humans with abundant unlabeled data to train deep neural networks. Classic methods on semi-supervised learning that have focused on transductive learning have not been fully exploited in the inductive framework followed by modern deep learning. The same holds for the manifold assumption-that similar examples should get the same prediction. In this work, we employ a transductive label propagation method that is based on the manifold assumption to make predictions on the entire dataset and use these predictions to generate pseudo-labels for the unlabeled data and train a deep neural network. At the core of the transductive method lies a nearest neighbor graph of the dataset that we create based on the embeddings of the same network. Therefore our learning process iterates between these two steps. We improve performance on several datasets especially in the few labels regime and show that our work is complementary to current state of the art.

Few-shot learning for deep neural networks is a highly challenging and key problem in many computer vision tasks. In this context, we are targeting knowledge transfer from a set with abundant data to other sets with few available examples. We propose two simple and effective solutions: (i) dense classification over feature maps, which for the first time studies local activations in the domain of few-shot learning, and (ii) implanting, that is, attaching new neurons to a previously trained network to learn new, task-specific features. Implanting enables training of multiple layers in the few-shot regime, departing from most related methods derived from metric learning that train only the final layer. Both contributions show consistent gains when used individually or jointly and we report state of the art performance on few-shot classification on miniImageNet.

We propose a novel method of deep spatial matching (DSM) for image retrieval. Initial ranking is based on image descriptors extracted from convolutional neural network activations by global pooling, as in recent state-of-the-art work. However, the same sparse 3D activation tensor is also approximated by a collection of local features. These local features are then robustly matched to approximate the optimal alignment of the tensors. This happens without any network modification, additional layers or training. No local feature detection happens on the original image. No local feature descriptors and no visual vocabulary are needed throughout the whole process. We experimentally show that the proposed method achieves the state-of-the-art performance on standard benchmarks across different network architectures and different global pooling methods. The highest gain in performance is achieved when diffusion on the nearest-neighbor graph of global descriptors is initiated from spatially verified images.

Despite the success of deep learning on representing images for particular object retrieval, recent studies show that the learned representations still lie on manifolds in a high dimensional space. This makes the Euclidean nearest neighbor search biased for this task. Exploring the manifolds online remains expensive even if a nearest neighbor graph has been computed offline. This work introduces an explicit embedding reducing manifold search to Euclidean search followed by dot product similarity search. This is equivalent to linear graph filtering of a sparse signal in the frequency domain. To speed up online search, we compute an approximate Fourier basis of the graph offline. We improve the state of art on particular object retrieval datasets including the challenging Instre dataset containing small objects. At a scale of 105 images, the offline cost is only a few hours, while query time is comparable to standard similarity search.

In this work we present a novel unsupervised framework for hard training example mining. The only input to the method is a collection of images relevant to the target application and a meaningful initial representation, provided e.g. by pre-trained CNN. Positive examples are distant points on a single manifold, while negative examples are nearby points on different manifolds. Both types of examples are revealed by disagreements between Euclidean and manifold similarities. The discovered examples can be used in training with any discriminative loss. The method is applied to unsupervised fine-tuning of pre-trained networks for fine-grained classification and particular object retrieval. Our models are on par or are outperforming prior models that are fully or partially supervised.

In this paper we address issues with image retrieval benchmarking on standard and popular Oxford 5k and Paris 6k datasets. In particular, annotation errors, the size of the dataset, and the level of challenge are addressed: new annotation for both datasets is created with an extra attention to the reliability of the ground truth. Three new protocols of varying difficulty are introduced. The protocols allow fair comparison between different methods, including those using a dataset pre-processing stage. For each dataset, 15 new challenging queries are introduced. Finally, a new set of 1M hard, semi-automatically cleaned distractors is selected. An extensive comparison of the state-of-the-art methods is performed on the new benchmark. Different types of methods are evaluated, ranging from local-feature-based to modern CNN based methods. The best results are achieved by taking the best of the two worlds. Most importantly, image retrieval appears far from being solved.

Severe background clutter is challenging in many computer vision tasks, including large-scale image retrieval. Global descriptors, that are popular due to their memory and search efficiency, are especially prone to corruption by such a clutter. Eliminating the impact of the clutter on the image descriptor increases the chance of retrieving relevant images and prevents topic drift due to actually retrieving the clutter in the case of query expansion. In this work, we propose a novel salient region detection method. It captures, in an unsupervised manner, patterns that are both discriminative and common in the dataset. Saliency is based on a centrality measure of a nearest neighbor graph constructed from regional CNN representations of dataset images. The descriptors derived from the salient regions improve particular object retrieval, most noticeably in a large collections containing small objects.

Query expansion is a popular method to improve the quality of image retrieval with both conventional and CNN representations. It has been so far limited to global image similarity. This work focuses on diffusion, a mechanism that captures the image manifold in the feature space. An efficient off-line stage allows optional reduction in the number of stored regions. In the on-line stage, the proposed handling of unseen queries in the indexing stage removes additional computation to adjust the precomputed data. We perform diffusion through a sparse linear system solver, yielding practical query times well below one second. Experimentally, we observe a significant boost in performance of image retrieval with compact CNN descriptors on standard benchmarks, especially when the query object covers only a small part of the image. Small objects have been a common failure case of CNN-based retrieval.

Part-based image classification aims at representing categories by small sets of learned discriminative parts, upon which an image representation is built. Considered as a promising avenue a decade ago, this direction has been neglected since the advent of deep neural networks. In this context, this paper brings two contributions: first, this work proceeds one step further compared to recent part-based models (PBM), focusing on how to learn parts without using any labeled data. Instead of learning a set of parts per class, as generally performed in the PBM literature, the proposed approach both constructs a partition of a given set of images into visually similar groups, and subsequently learns a set of discriminative parts per group in a fully unsupervised fashion. This strategy opens the door to the use of PBM in new applications where labeled data are typically not available, such as instance-based image retrieval. Second, this paper shows that despite the recent success of end-to-end models, explicit part learning can still boost classification performance. We experimentally show that our learned parts can help building efficient image representations, which outperform state-of-the art Deep Convolutional Neural Networks (DCNN) on both classification and retrieval tasks.

International audience

Recent works show that image comparison based on local descriptors is corrupted by visual bursts, which tend to dominate the image similarity. The existing strategies, like power-law normalization, improve the results by discounting the contribution of visual bursts to the image similarity.

Large scale duplicate detection, clustering and mining of documents or images has been conventionally treated with seed detection via hashing, followed by seed growing heuristics using fast search. Principled clustering methods, especially kernelized and spectral ones, have higher complexity and are difficult to scale above millions. Under the assumption of documents or images embedded in Euclidean space, we revisit recent advances in approximate k-means variants, and borrow their best ingredients to introduce a new one, inverted-quantized k-means (IQ-means). Key underlying concepts are quantization of data points and multi-index based inverted search from centroids to cells. Its quantization is a form of hashing and analogous to seed detection, while its updates are analogous to seed growing, yet principled in the sense of distortion minimization. We further design a dynamic variant that is able to determine the number of clusters k in a single run at nearly zero additional cost. Combined with powerful deep learned representations, we achieve clustering of a 100 million image collection on a single machine in less than one hour.

We present a simple vector quantizer that combines low distortion with fast search and apply it to approximate nearest neighbor (ANN) search in high dimensional spaces. Leveraging the very same data structure that is used to provide non-exhaustive search, i.e., inverted lists or a multi-index, the idea is to locally optimize an individual product quantizer (PQ) per cell and use it to encode residuals. Local optimization is over rotation and space decomposition, interestingly, we apply a parametric solution that assumes a normal distribution and is extremely fast to train. With a reasonable space and time overhead that is constant in the data size, we set a new state-of-the-art on several public datasets, including a billion-scale one.

Inspired by the close relation between nearest neighbor search and clustering in high-dimensional spaces as well as the success of one helping to solve the other, we introduce a new paradigm where both problems are solved simultaneously. Our solution is recursive, not in the size of input data but in the number of dimensions. One result is a clustering algorithm that is tuned to small codebooks but does not need all data in memory at the same time and is practically constant in the data size. As a by-product, a tree structure performs either exact or approximate quantization on trained centroids, the latter being not very precise but extremely fast. A lesser contribution is a new indexing scheme for image retrieval that exploits multiple small codebooks to provide an arbitrarily fine partition of the descriptor space. Large scale experiments on public datasets exhibit state of the art performance and remarkable generalization.

This paper considers a family of metrics to compare images based on their local descriptors. It encompasses the VLAD descriptor and matching techniques such as Hamming Embedding. Making the bridge between these approaches leads us to propose a match kernel that takes the best of existing techniques by combining an aggregation procedure with a selective match kernel. Finally, the representation underpinning this kernel is approximated, providing a large scale image search both precise and scalable, as shown by our experiments on several benchmarks.

Multimodal streams of sensory information are naturally parsed and integrated by humans using signal-level feature extraction and higher level cognitive processes. Detection of attention-invoking audiovisual segments is formulated in this work on the basis of saliency models for the audio, visual, and textual information conveyed in a video stream. Aural or auditory saliency is assessed by cues that quantify multifrequency waveform modulations, extracted through nonlinear operators and energy tracking. Visual saliency is measured through a spatiotemporal attention model driven by intensity, color, and orientation. Textual or linguistic saliency is extracted from part-of-speech tagging on the subtitles information available with most movie distributions. The individual saliency streams, obtained from modality-depended cues, are integrated in a multimodal saliency curve, modeling the time-varying perceptual importance of the composite video stream and signifying prevailing sensory events. The multimodal saliency representation forms the basis of a generic, bottom-up video summarization algorithm. Different fusion schemes are evaluated on a movie database of multimodal saliency annotations with comparative results provided across modalities. The produced summaries, based on low-level features and content-independent fusion and selection, are of subjectively high aesthetic and informative quality.

Many problems, including feature selection, vocabulary learning, location and landmark recognition, structure from motion and 3d reconstruction, rely on a learning process that involves wide-baseline matching on multiple views of the same object or scene. In practical large scale image retrieval applications however, most images depict unique views where this idea does not apply. We exploit self-similarities, symmetries and repeating patterns to select features within a single image. We achieve the same performance compared to the full feature set with only a small fraction of its index size on a dataset of unique views of buildings or urban scenes, in the presence of one million distractors of similar nature. Our best solution is linear in the number of correspondences, with practical running times of just a few milliseconds.

We present a local feature detector that is able to detect regions of arbitrary scale and shape, without scale space construction. We compute a weighted distance map on image gradient, using our exact linear-time algorithm, a variant of group marching for Euclidean space. We find the weighted medial axis by extending residues, typically used in Voronoi skeletons. We decompose the medial axis into a graph representing image structure in terms of peaks and saddle points. A duality property enables reconstruction of regions using the same marching method. We greedily group regions taking both contrast and shape into account. On the way, we select regions according to our shape fragmentation factor, favoring those well enclosed by boundaries-even incomplete. We achieve state of the art performance in matching and retrieval experiments with reduced memory and computational requirements.

A wide range of properties and assumptions determine the most appropriate spatial matching model for an application, e.g. recognition, detection, registration, or large scale image retrieval. Most notably, these include discriminative power, geometric invariance, rigidity constraints, mapping constraints, assumptions made on the underlying features or descriptors and, of course, computational complexity. Having image retrieval in mind, we present a very simple model inspired by Hough voting in the transformation space, where votes arise from single feature correspondences. A relaxed matching process allows for multiple matching surfaces or non-rigid objects under one-to-one mapping, yet is linear in the number of correspondences. We apply it to geometry re-ranking in a search engine, yielding superior performance with the same space requirements but a dramatic speed-up compared to the state of the art.

We propose a scalable logo recognition approach that extends the common bag-of-words model and incorporates local geometry in the indexing process. Given a query image and a large logo database, the goal is to recognize the logo contained in the query, if any. We locally group features in triples using multi-scale Delaunay triangulation and represent triangles by signatures capturing both visual appearance and local geometry. Each class is represented by the union of such signatures over all instances in the class. We see large scale recognition as a sub-linear search problem where signatures of the query image are looked up in an inverted index structure of the class models. We evaluate our approach on a large-scale logo recognition dataset with more than four thousand classes.

State of the art data mining and image retrieval in community photo collections typically focus on popular subsets, e.g. images containing landmarks or associated to Wikipedia articles. We propose an image clustering scheme that, seen as vector quantization compresses a large corpus of images by grouping visually consistent ones while providing a guaranteed distortion bound. This allows us, for instance, to represent the visual content of all thousands of images depicting the Parthenon in just a few dozens of scene maps and still be able to retrieve any single, isolated, non-landmark image like a house or graffiti on a wall. Starting from a geo-tagged dataset, we first group images geographically and then visually, where each visual cluster is assumed to depict different views of the the same scene. We align all views to one reference image and construct a 2D scene map by preserving details from all images while discarding repeating visual features. Our indexing, retrieval and spatial matching scheme then operates directly on scene maps. We evaluate the precision of the proposed method on a challenging one-million urban image dataset.

We present a new approach to image indexing and retrieval, which integrates appearance with global image geometry in the indexing process, while enjoying robustness against viewpoint change, photometric variations, occlusion, and background clutter. We exploit shape parameters of local features to estimate image alignment via a single correspondence. Then, for each feature, we construct a sparse spatial map of all remaining features, encoding their normalized position and appearance, typically vector quantized to visual word. An image is represented by a collection of such feature maps and RANSAC-like matching is reduced to a number of set intersections. Because the induced dissimilarity is still not a metric, we extend min-wise independent permutations to collections of sets and derive a similarity measure for feature map collections. We then exploit sparseness to build an inverted file whereby the retrieval process is sub-linear in the total number of images, ideally linear in the number of relevant ones. We achieve excellent performance on 10^4 images, with a query time in the order of milliseconds.

Several spatiotemporal feature point detectors have been used in video analysis for action recognition. Feature points are detected using a number of measures, namely saliency, cornerness, periodicity, motion activity etc. Each of these measures is usually intensity-based and provides a different trade-off between density and informativeness. In this paper, we use saliency for feature point detection in videos and incorporate color and motion apart from intensity. Our method uses a multi-scale volumetric representation of the video and involves spatiotemporal operations at the voxel level. Saliency is computed by a global minimization process constrained by pure volumetric constraints, each of them being related to an informative visual aspect, namely spatial proximity, scale and feature similarity (intensity, color, motion). Points are selected as the extrema of the saliency response and prove to balance well between density and informativeness. We provide an intuitive view of the detected points and visual comparisons against state-of-the-art space-time detectors. Our detector outperforms them on the KTH dataset using nearest-neighbor classifiers and ranks among the top using different classification frameworks. Statistics and comparisons are also performed on the more difficult Hollywood human actions (HOHA) dataset increasing the performance compared to current published results.

Detection of perceptually important video events is formulated here on the basis of saliency models for the audio, visual and textual information conveyed in a video stream. Audio saliency is assessed by cues that quantify multifrequency waveform modulations, extracted through nonlinear operators and energy tracking. Visual saliency is measured through a spatiotemporal attention model driven by intensity, color and motion. Text saliency is extracted from part-of-speech tagging on the subtitles information available with most movie distributions. The various modality curves are integrated in a single attention curve, where the presence of an event may be signified in one or multiple domains. This multimodal saliency curve is the basis of a bottom-up video summarization algorithm, that refines results from unimodal or audiovisual-based skimming. The algorithm performs favorably for video summarization in terms of informativeness and enjoyability.

In this paper we investigate detection of high-level concepts in multimedia content through an integrated approach of visual thesaurus analysis and visual context. In the former, detection is based on model vectors that represent image composition in terms of region types, obtained through clustering over a large data set. The latter deals with two aspects, namely high-level concepts and region types of the thesaurus, employing a model of a priori specified semantic relations among concepts and automatically extracted topological relations among region types; thus it combines both conceptual and topological context. A set of algorithms is presented, which modify either the confidence values of detected concepts, or the model vectors based on which detection is performed. Visual context exploitation is evaluated on TRECVID and Corel data sets and compared to a number of related visual thesaurus approaches.

Based on perceptual and computational attention modeling studies, we formulate measures of saliency for an audiovisual stream. Audio saliency is captured by signal modulations and related multi-frequency band features, extracted through nonlinear operators and energy tracking. Visual saliency is measured by means of a spatiotemporal attention model driven by various feature cues (intensity, color, motion). Audio and video curves are integrated in a single attention curve, where events may be enhanced, suppressed or vanished. The presence of salient events is signified on this audiovisual curve by geometrical features such as local extrema, sharp transition points and level sets. An audiovisual saliency-based movie summarization algorithm is proposed and evaluated. The algorithm is shown to perform very well in terms of summary informativeness and enjoyability for movie clips of various genres.

In this paper we focus on scene classification and detection of high-level concepts within multimedia documents, by introducing an intermediate contextual approach as a means of exploiting the visual context of images. More specifically, we introduce and model a novel relational knowledge representation, founded on topological and semantic relations between the concepts of an image. We further develop an algorithm to address computationally efficient handling of visual context and extraction of mid-level region characteristics. Based on the proposed knowledge model, we combine the notion of visual context with region semantics, in order to exploit their efficacy in dealing with scene classification problems. Finally, initial experimental results are presented, in order to demonstrate possible applications of the proposed methodology.

In this paper, we propose a framework to extend semantic labeling of images to video shot sequences and achieve efficient and semantic-aware spatiotemporal video segmentation. This task faces two major challenges, namely the temporal variations within a video sequence which affect image segmentation and labeling, and the computational cost of region labeling. Guided by these limitations, we design a method where spatiotemporal segmentation and object labeling are coupled to achieve semantic annotation of video shots. An internal graph structure that describes both visual and semantic properties of image and video regions is adopted. The process of spatiotemporal semantic segmentation is subdivided in two stages: Firstly, the video shot is split into small block of frames. Spatiotemporal regions (volumes) are extracted and labeled individually within each block. Then, we iteratively merge consecutive blocks by a matching procedure which considers both semantic and visual properties. Results on real video sequences show the potential of our approach.

This paper describessalienShrink, a method to denoise images based on computing a map of salient coefficients in the wavelet domain and use it to improve common denoising algorithms. By salient, we refer to those coefficients that correspond mostly to pure signal and should therefore be preserved throughout the denoising procedure. We use a computationally efficient model to detect salient regions in the bands of the multiresolution wavelet transform. These regions are used to obtain a more accurate estimate of the noise level, improving the performance of existing well known shrinkage methods. Extensive experimental results on theBiShrinkmethod show that the proposed method effectively enhances PSNR and improves the visual quality of the denoised images.

A saliency-based method for generating video summaries is presented, which exploits coupled audiovisual information from both media streams. Efficient and advanced speech and image processing algorithms to detect key frames that are acoustically and visually salient are used. Promising results are shown from experiments on a movie database.

In this paper, we present a framework for simultaneous image segmentation and object labeling leading to automatic image annotation. Focusing on semantic analysis of images, it contributes to knowledge-assisted multimedia analysis and bridging the gap between semantics and low level visual features. The proposed framework operates at semantic level using possible semantic labels, formally represented as fuzzy sets, to make decisions on handling image regions instead of visual features used traditionally. In order to stress its independence of a specific image segmentation approach we have modified two well known region growing algorithms, i.e., watershed and recursive shortest spanning tree, and compared them to their traditional counterparts. Additionally, a visual context representation and analysis approach is presented, blending global knowledge in interpreting each object locally. Contextual information is based on a novel semantic processing methodology, employing fuzzy algebra and ontological taxonomic knowledge representation. In this process, utilization of contextual knowledge re-adjusts labeling results of semantic region growing, by means of fine-tuning membership degrees of detected concepts. The performance of the overall methodology is evaluated on a real-life still image dataset from two popular domains

Personalized content retrieval aims at improving the retrieval process by taking into account the particular interests of individual users. However, not all user preferences are relevant in all situations. It is well known that human preferences are complex, multiple, heterogeneous, changing, even contradictory, and should be understood in context with the user goals and tasks at hand. In this paper, we propose a method to build a dynamic representation of the semantic context of ongoing retrieval tasks, which is used to activate different subsets of user interests at runtime, in a way that out-of-context preferences are discarded. Our approach is based on an ontology-driven representation of the domain of discourse, providing enriched descriptions of the semantics involved in retrieval actions and preferences, and enabling the definition of effective means to relate preferences and context

In this article, an approach to semantic image analysis is presented. Under the proposed approach, ontologies are used to capture general, spatial, and contextual knowledge of a domain, and a genetic algorithm is applied to realize the final annotation. The employed domain knowledge considers high-level information in terms of the concepts of interest of the examined domain, contextual information in the form of fuzzy ontological relations, as well as low-level information in terms of prototypical low-level visual descriptors. To account for the inherent ambiguity in visual information, uncertainty has been introduced in the spatial relations definition. First, an initial hypothesis set of graded annotations is produced for each image region, and then context is exploited to update appropriately the estimated degrees of confidence. Finally, a genetic algorithm is applied to decide the most plausible annotation by utilizing the visual and the spatial concepts definitions included in the domain ontology. Experiments with a collection of photographs belonging to two different domains demonstrate the performance of the proposed approach.

In the last few years, numerous multimedia archives have made extensive use of digitized storage and annotation technologies. Still, the development of single points of access, providing common and uniform access to their data, despite the efforts and accomplishments of standardization organizations, has remained an open issue as it involves the integration of various large-scale heterogeneous and heterolingual systems. This paper describes a mediator system that achieves architectural integration through an extended three-tier architecture and content integration through semantic modeling. The described system has successfully integrated five multimedia archives, quite different in nature and content from each other, while also providing easy and scalable inclusion of more archives in the future.

In natural vision, we center our fixation on the most informative points in a scene in order to reduce our overall uncertainty about the scene and help interpret it. Even if we are looking for a specific stimulus around us, we face a great amount of uncertainty since that stimulus could be in any spatial location. Visual attention (VA) schemes have been proposed by researchers to account for the ability of the human eye to quickly fixate on informative regions. Recently, VA in images, and especially saliency-based VA, became an active research topic of the computer vision community. The proposed work provides an extension towards VA in video sequences by integrating spatiotemporal information. The potential applications include video classification, scene understanding, surveillance and segmentation

We follow a fuzzy relational approach to knowledge representation. With the use of semantic fuzzy relations we define and extract the semantic context out of a set of semantic entities. Based on this, we then proceed to the case of information retrieval and explain how the three participating contexts, namely the context of the query, the context of the document and the context of the user, can be estimated and utilized towards the achievement of more intuitive information services.

Several visual attention (VA) schemes have been proposed with the saliency-based ones being the most popular. The proposed work provides an extension towards VA in video sequences by incorporating the temporal dimension. The architecture is presented in detail and potential applications are investigated. We expect that the extended VA scheme can reveal interesting events across the sequence like occlusions and short occurrences of objects, providing a basis for video surveillance (e.g. intruder detection), segmentation, and summarization applications.

Face detection is becoming an important tool in the framework of many multimedia applications. Several face detection algorithms based on skin color characteristics have appeared in the literature. Most of them have generalization problems due to the skin color model they use. We present a study which attempts to minimize the generalization problem by combining the M-RSST color segmentation algorithm with a Gaussian model of the skin color distribution and global shape features. Moreover by associating the resultant segments with a face probability we can index and retrieve facial images from multimedia databases

A novel method for two-dimensional curve normalization with respect to affine transformations is presented in this paper, allowing an affine-invariant curve representation to be obtained without any actual loss of information on the original curve. It can be applied as a pre-processing step to any shape representation, classification, recognition or retrieval technique, since it effectively decouples the problem of affine-invariant description from feature extraction and pattern matching. Curves estimated from object contours are first modeled by cubic B-splines and then normalized in several steps in order to eliminate translation, scaling, skew, starting point, rotation and reflection transformations, based on a combination of curve features including moments and Fourier descriptors.

An efficient technique for summarization of stereoscopic video sequences is presented, which extracts a small but meaningful set of video frames using a content-based sampling algorithm. The proposed video-content representation provides the capability of browsing digital stereoscopic video sequences and performing more efficient content-based queries and indexing. Each stereoscopic video sequence is first partitioned into shots by applying a shot-cut detection algorithm so that frames (or stereo pairs) of similar visual characteristics are gathered together. Each shot is then analyzed using stereo-imaging techniques, and the disparity field, occluded areas, and depth map are estimated. A multiresolution implementation of the recursive shortest spanning tree (RSST) algorithm is applied for color and depth segmentation, while fusion of color and depth segments is employed for reliable video object extraction. In particular, color segments are projected onto depth segments so that video objects on the same depth plane are retained, while at the same time accurate object boundaries are extracted. Feature vectors are then constructed using multidimensional fuzzy classification of segment features including size, location, color, and depth. Shot selection is accomplished by clustering similar shots based on the generalized Lloyd-Max algorithm, while for a given shot, key frames are extracted using an optimization method for locating frames of minimally correlated feature vectors. For efficient implementation of the latter method, a genetic algorithm is used. Experimental results are presented, which indicate the reliable performance of the proposed scheme on real-life stereoscopic video sequences.

A framework for video content representation is proposed in this paper for extracting limited, but meaningful, information of video data directly from MPEG compressed domain. First, the traditional frame-based representation is transformed to a feature-based one. Then, all features are gathered together using a fuzzy formulation and extraction of several key frames is performed for each shot in a content-based rate sampling framework. In particular, our approach is based on minimization of a cross-correlation criterion among video frames of a given shot so as to be located a set of minimally correlated feature vectors. Experimental results indicating the good performance of the proposed scheme are also presented.

An efficient video content representation is proposed using optimal extraction of characteristic frames and scenes. This representation, apart from providing browsing capabilities to digital video databases, also allows more efficient content-based queries and indexing. For performing the frame/scene extraction, a feature vector formulation of the images is proposed based on color and motion segmentation. Then, the scene selection is accomplished by clustering similar scenes based on a distortion criterion. Frame selection is performed using an optimization method for locating a set of minimally correlated feature vectors.