This paper describes GeoPl@ntNet, an interactive web application designed to make Essential Biodiversity Variables accessible and understandable to everyone through dynamic maps and fact sheets. Its core purpose is to allow users to explore high-resolution AI-generated maps of species distributions, habitat types, and biodiversity indicators across Europe. These maps, developed through a cascading pipeline involving convolutional neural networks and large language models, provide an intuitive yet information-rich interface to better understand biodiversity, with resolutions as precise as 50×50 meters. The website also enables exploration of specific regions, allowing users to select areas of interest on the map (e.g., urban green spaces, protected areas, or riverbanks) to view local species and their coverage. Additionally, GeoPl@ntNet generates comprehensive reports for selected regions, including insights into the number of protected species, invasive species, and endemic species.

Deep learning models have emerged as a promising alternative to conventional approaches for plant disease identification, a critical challenge in agricultural production. However, the existing plant disease datasets are insufficient to address the complexities of real-world agricultural scenarios, such as multi crop disease, unseen, few-shot, and domain shift adaptation. Additionally, the lack of standardized evaluation protocols and benchmark datasets hinders the fair evaluation of models against these challenges. To bridge this gap, we introduce Deep-Plant-Disease, the largest and most diverse dataset with novel text data designed to enhance model generalization in multi crop disease identification. We revisit and reformulate the task by establishing a standardized evaluation framework that supports consistent benchmarking and guides future research. Through experiments, we further validate the robustness and adaptability of models trained on our dataset, highlighting their effective transferability to real-world agricultural challenges.

Average- K classification is an alternative to top- K clas-sification in which the number of labels returned varies with the ambiguity of the input image but must average to K over all the samples. A simple method to solve this task is to threshold the softmax output of a model trained with the cross-entropy loss. This approach is theoretically proven to be asymptotically consistent, but it is not guaranteed to be optimal for a finite set of samples. In this paper, we pro-pose a new loss function based on a multi-label classification head in addition to the classical softmax. This second head is trained using pseudo-labels generated by thresholding the softmax head while guaranteeing that K classes are returned on average. We show that this approach allows the model to better capture ambiguities between classes and, as a result, to return more consistent sets of possible classes. Experiments on two datasets from the literature demon-strate that our approach outperforms the softmax baseline, as well as several other loss functions more generally de-signed for weakly supervised multi-label classification. The gains are larger the higher the uncertainty, especially for classes with few samples.

The difficulty of monitoring biodiversity at fine scales and over large areas limits ecological knowledge and conservation efforts. To fill this gap, Species Distribution Models (SDMs) predict species across space from spatially explicit features. Yet, they face the challenge of integrating the rich but heterogeneous data made available over the past decade, notably millions of opportunistic species observations and standardized surveys, as well as multi-modal remote sensing data.In light of that, we have designed and developed a new European-scale dataset for SDMs at high spatial resolution (10--50m), including more than 10k species (i.e., most of the European flora). The dataset comprises 5M heterogeneous Presence-Only records and 90k exhaustive Presence-Absence survey records, all accompanied by diverse environmental rasters (e.g., elevation, human footprint, and soil) traditionally used in SDMs. In addition, it provides Sentinel-2 RGB and NIR satellite images with 10 m resolution, a 20-year time series of climatic variables, and satellite time series from the Landsat program.In addition to the data, we provide an openly accessible SDM benchmark (hosted on Kaggle), which has already attracted an active community and a set of strong baselines for single predictor/modality and multimodal approaches.All resources, e.g., the dataset, pre-trained models, and baseline methods (in the form of notebooks), are available on Kaggle, allowing one to start with our dataset literally with two mouse clicks.

Deep learning models have demonstrated great promise in plant disease identification. However, existing approaches often face challenges when dealing with unseen crop-disease pairs, limiting their practicality in real-world settings. This research addresses the gap between known and unknown (unseen) plant disease identification. Our study pioneers the exploration of the zero-shot setting within this domain, offering a new perspective to conceptualizing plant disease identification. Specifically, we introduce the novel Cross Learning Vision Transformer (CL-ViT) model, incorporating self-supervised learning, in contrast to the previous state-of-the-art, FF-ViT, which emphasizes conceptual feature disentanglement with a synthetic feature generation framework. Through comprehensive analyses, we demonstrate that our novel model outperforms state-of-the-art models in both accuracy performance and visualization analysis. This study establishes a new benchmark and marks a significant advancement in the field of plant disease identification, paving the way for more robust and efficient plant disease identification systems . The code is available at https://github.com/abelchai/Cross-Learning-Vision-Transformer-CL-ViT .

Training in Feed Forward Deep Neural Networks is a memory-intensive operation which is usually performed on GPUs with limited memory capacities. This may force data scientists to limit the depth of the models or the resolution of the input data if data does not fit in the GPU memory. The re-materialization technique, whose idea comes from the checkpointing strategies developed in the Automatic Differentiation literature, allows data scientists to limit the memory requirements related to the storage of intermediate data (activations), at the cost of an increase in the computational cost. This paper introduces a new strategy of re-materialization of activations that significantly reduces memory usage. It consists in selecting which activations are saved and which activations are deleted during the forward phase and then recomputing the deleted activations when they are needed during the backward phase. We propose an original computation model that combines two types of activation savings: either only storing the layer inputs or recording the complete history of operations that produced the outputs. This paper focuses on the fully heterogeneous case, where the computation time and the memory requirement of each layer is different. We prove that finding the optimal solution is NP-hard and that classical techniques from Automatic Differentiation literature do not apply. Moreover, the classical assumption of memory persistence of materialized activations, used to simplify the search of optimal solutions, does not hold anymore. Thus, we propose a weak memory persistence property and provide a dynamic program to compute the optimal sequence of computations. This algorithm is made available through the Rotor software, a PyTorch plug-in dealing with any network consisting of a sequence of layers, each of them having an arbitrarily complex structure. Through extensive experiments, we show that our implementation consistently outperforms existing re-materialization approaches for a large class of networks, image sizes, and batch sizes.

With the advent of Deep Learning, people have begun to use it with computer vision approaches to identify plant diseases on a large scale targeting multiple crops and diseases. However, this requires a large amount of plant disease data, which is often not readily available, and the cost of acquiring disease images is high. Thus, developing a generalized model for recognizing unseen classes is very important and remains a major challenge to date. Existing methods solve the problem with general supervised recognition tasks based on the seen composition of the crop and the disease. However, ignoring the composition of unseen classes during model training can lead to a reduction in model generalisation. Therefore, in this work, we propose a new approach that leverages the visual features of crop and disease from the seen composition, using them to learn the features of unseen crop-disease composition classes. We show that our proposed method can improve the classification performance of these unseen classes and outperform the state-of-the-art in the identification of multiple crop-diseases.

In modern classification tasks, the number of labels is getting larger and larger, as is the size of the datasets encountered in practice. As the number of classes increases, class ambiguity and class imbalance become more and more problematic to achieve high top-1 accuracy. Meanwhile, Top-K metrics (metrics allowing K guesses) have become popular, especially for performance reporting. Yet, proposing top-K losses tailored for deep learning remains a challenge, both theoretically and practically. In this paper we introduce a stochastic top-K hinge loss inspired by recent developments on top-K calibrated losses. Our proposal is based on the smoothing of the top-K operator building on the flexible "perturbed optimizer" framework. We show that our loss function performs very well in the case of balanced datasets, while benefiting from a significantly lower computational time than the state-of-the-art top-K loss function. In addition, we propose a simple variant of our loss for the imbalanced case. Experiments on a heavy-tailed dataset show that our loss function significantly outperforms other baseline loss functions.

In more and more application areas, we are witnessing the emergence of complex workflows that combine computing, analytics and learning. They often require a hybrid execution infrastructure with IoT devices interconnected to cloud/HPC systems (aka Computing Continuum). Such workflows are subject to complex constraints and requirements in terms of performance, resource usage, energy consumption and financial costs. This makes it challenging to optimize their configuration and deployment. We propose a methodology to support the optimization of real-life applications on the Edge-to-Cloud Continuum. We implement it as an extension of E2Clab, a previously proposed framework supporting the complete experimental cycle across the Edge-to-Cloud Continuum. Our approach relies on a rigorous analysis of possible configurations in a controlled testbed environment to understand their behaviour and related performance tradeoffs. We illustrate our methodology by optimizing Pl@ntNet, a world-wide plant identification application. Our methodology can be generalized to other applications in the Edge-to-Cloud Continuum.

Convolutional Neural Networks (CNNs) are statistical models suited for learning complex visual patterns. In the context of Species Distribution Models (SDM) and in line with predictions of landscape ecology and island biogeography, CNN could grasp how local landscape structure affects prediction of species occurrence in SDMs. The prediction can thus reflect the signatures of entangled ecological processes. Although previous machine-learning based SDMs can learn complex influences of environmental predictors, they cannot acknowledge the influence of environmental structure in local landscapes (hence denoted "punctual models"). In this study, we applied CNNs to a large dataset of plant occurrences in France (GBIF), on a large taxonomical scale, to predict ranked relative probability of species (by joint learning) to any geographical position. We examined the way local environmental landscapes improve prediction by performing alternative CNN models deprived of information on landscape heterogeneity and structure ("ablation experiments"). We found that the landscape structure around location crucially contributed to improve predictive performance of CNN-SDMs. CNN models can classify the predicted distributions of many species, as other joint modelling approaches, but they further prove efficient in identifying the influence of local environmental landscapes. CNN can then represent signatures of spatially structured environmental drivers. The prediction gain is noticeable for rare species, which open promising perspectives for biodiversity monitoring and conservation strategies. Therefore, the approach is of both theoretical and practical interest. We discuss the way to test hypotheses on the patterns learnt by CNN, which should be essential for further interpretation of the ecological processes at play.

Several recent studies have proposed an automatic plant disease identification system based on deep learning. However, these approaches are generally based on learned classification models with target classes of joint host species-disease pairs that may not allow optimal use of the available information. This is because these approaches require distinguishing between similar host species or diseases, and more importantly, they have limited scalability due to the size of a network gradually increases as new classes are added, despite the fact that information on host species or diseases is already available. This constraint is all the more important as it can be difficult to collect/establish a specific list of all diseases for each host plant species in an actual application. In this paper, we address the problems by proposing a new conditional multi-task learning (CMTL) approach which allows the distribution of host species and disease characteristics learned simultaneously with a conditional link between them. This conditioning is formed in such a way that the knowledge to infer the prediction of one concept (the diseases) depends on the other concept (the host species), which corresponds to the way plant pathologists used to infer the diseases of the host species. We show that our approach can improve the performance of plant disease identification compared to the usual species-disease pair modeling in the previous studies. Meanwhile, we also compose a new dataset on plant disease identification that could serve as an important benchmark in this field.

[email protected] is a world-scale participatory platform and information system dedicated to the monitoring of plant biodiversity through image-based plant identification. Nowadays, the mobile front-end of [email protected] has been downloaded by more than 4 millions users in about 170 countries and an active community of contributors produce and revise new observations everyday (it is often referred to as the "shazam of plants"). This paper presents a business proposal allowing enterprises or organizations to set up their own private collaborative workflow within [email protected] information system. The main added value is to allow them working on their own business object (e.g. plant disease diagnostic, deficiency measurements, railway lines maintenance, fishery surveillance, etc.) and with their own community of contributors and end-users (employees, sales representatives, clients, observers network, etc.). This business idea answers to a growing demand in agriculture and environmental economics. Actors in these domains begin to know that machine learning techniques are mature enough but the lack of training data and of efficient tools to collect them is a major breakthrough. A collaborative platform like [email protected] is the ideal tool to bridge this gap. It initiates a powerful positive feedback loop boosting the production of training data while improving the work of the employees.

Large-scale annotated corpora have yielded impressive performance improvements in computer vision and multimedia content analysis. However, such datasets depend on an enormous amount of human labeling effort. When the labels correspond to well-known concepts, it is straightforward to train the annotators by giving a few examples with known answers. It is also straightforward to judge the quality of their labels. Neither is true when there are thousands of complex domain-specific labels. Training on all labels is infeasible and the quality of an annotator's judgements may be vastly different for some subsets of labels than for others. This paper proposes a set of data-driven algorithms to 1) train image annotators on how to disambiguate among automatically generated candidate labels, 2) evaluate the quality of annotators' label suggestions, and 3) weigh predictions. The algorithms adapt to the skills of each annotator both in the questions asked and the weights given to their answers. The underlying judgements are Bayesian, based on adaptive priors. We measure the benefits of these algorithms on a live user experiment related to image-based plant identification involving around 1000 people. The proposed methods are shown to enable huge gains in annotation accuracy. A standard user can correctly label around 2% of our data. This goes up to 80% with machine learning assisted training and assignment and up to almost 90% when doing a weighted combination of several annotators' labels.

This paper addresses the challenge of devising new representation learning algorithms that overcome the lack of interpretability of classical visual models. Therefore, it introduces a new recursive visual patch selection technique built on top of a Shared Nearest Neighbors embedding method. The main contribution of the paper is to drastically reduce the high-dimensionality of such over-complete representation thanks to a recursive feature elimination method. We show that the number of spatial atoms of the representation can be reduced by up to two orders of magnitude without much degrading the encoded information. The resulting representations are shown to provide competitive image classification performance with the state-of-the-art while enabling to learn highly interpretable visual models.

Large scale biodiversity monitoring is essential for sustainable development (earth stewardship). With the recent advances in computer vision, we see the emergence of more and more effective identification tools allowing to set-up large-scale data collection platforms such as the popular [email protected] initiative that allow to reuse interaction data. Although it covers only a fraction of the world flora, this platform is already being used by more than 300K people who produce tens of thousands of validated plant observations each year. This explicitly shared and validated data is only the tip of the iceberg. The real potential relies on the millions of raw image queries submitted by the users of the mobile application for which there is no human validation. People make such requests to get information on a plant along a hike or something they find in their garden but not know anything about. Allowing the exploitation of such contents in a fully automatic way could scale up the world-wide collection of implicit plant observations by several orders of magnitude, which can complement the explicit monitoring efforts. In this paper, we first survey existing automated plant identification systems through a five-year synthesis of the PlantCLEF benchmark and an impact study of the [email protected] platform. We then focus on the implicit monitoring scenario and discuss related research challenges at the frontier of computer science and biodiversity studies. Finally, we discuss the results of a preliminary study focused on implicit monitoring of invasive species in mobile search logs. We show that the results are promising but that there is room for improvement before being able to automatically share implicit observations within international platforms.

In a typical citizen science/crowdsourcing environment, the contributors label items. When there are few labels, it is straightforward to train contributors and judge the quality of their labels by giving a few examples with known answers. Neither is true when there are thousands of domain-specific labels and annotators with heterogeneous skills. This demo paper presents an Active User Training framework implemented as a serious game called ThePlantGame. It is based on a set of data-driven algorithms allowing to (i) actively train annotators, and (ii) evaluate the quality of contributors' answers on new test items to optimize predictions.

This paper presents an automatic system able to identify visual named-entities appearing in images and videos, among a list of 25,000 entities, aggregated from Wikipedia lists, and more specific websites. DigInPix is a generic application designed to identify different kinds of entities. In this first attempt, we only focus on logo identification (more generally on legal persons). The identification process mainly relies on an efficient CBIR system, searching in an indexed image database composed of 600,000 weak-labelled images crawled from Google Images. DigInPix proposes a responsive-design html5 interface [1] for testing purposes.

This paper introduces a new image representation relying on the spatial pooling of geometrically consistent visual matches. We therefore introduce a new match kernel based on the inverse rank of the shared nearest neighbors combined with local geometric constraints. To avoid overfitting and reduce processing costs, the dimensionality of the resulting over-complete representation is further reduced by hierarchically pooling the raw consistent matches according to their spatial position in the training images. The final image representation is obtained by concatenating the resulting feature vectors at several resolutions. Learning from these representations using a logistic regression classifier is shown to provide excellent fine-grained classification performances outperforming the results reported in the literature on several classification tasks.

This paper presents several improvements of [email protected], an image sharing and retrieval application for identifying plants [6]: (i) ported to most android platforms (ii) three times more data (iii) exploiting metadata as well as visual content in the identification process (iv) a new multi-plant-organ, multi-image and multi-feature merging strategy with separate indexes for each visual feature (v) integrating cross-languages functions. This paper also presents the new results achieved by our system in the ImageCLEF 2013 plant identification task and in real-world user trials.

This paper presents an interactive information retrieval scheme extending classical faceted search with fine-grained visual facets. Consistent visual words are extracted offline by constructing a matching graph at the region image level and splitting the graph in relevant clusters. Visual and textual facets are then indexed in the same faceted search engine, all of them being dynamically reranked according to their filtering efficiency before being presented to the end-user. To enhance their interpretability, each facet value is illustrated by a representative picture automatically selected from the matching graph in an adaptive way.

This paper considers the problem of recognizing legal entities in visual contents in a similar way to named-entity recognizers for text documents. Whereas previous works were restricted to the recognition of a few tens of logotypes, we generalize the problem to the recognition of thousands of legal persons, each being modeled by a rich corporate identity automatically built from web images. We introduce a new geometrically-consistent instance-based classification method that is shown to outperform state-of-the-art techniques on several challenging datasets while being much more scalable. Further experiments performed on an automatic web crawl of 5,824 legal entities demonstrates the scalability of the approach.

[email protected] is an image sharing and retrieval application for the identification of plants, available on iPhone and iPad devices. Contrary to previous content-based identification applications it can work with several parts of the plant including flowers, leaves, fruits and bark. It also allows integrating user's observations in the database thanks to a collaborative workflow involving the members of a social network specialized on plants. Data collected so far makes it one of the largest mobile plant identification tool.

Who said What, Where and How? How are images, video and stories spreading out? Who produces the information? OTMedia addresses these questions by collecting, enriching and analysing continuously more than 1500 streams of French media from TV Radio, Web, AFP, and Twitter. Two studies on media produced by end users with the OTMedia framework are presented.

State-of-the-art visual search methods allow retrieving efficiently small rigid objects in very large image datasets (e.g. logos, paintings, etc.). User's perception of the classical query-by-window paradigm is however affected by the fact that many submitted queries actually return nothing or only junk results. We demonstrate in this demo that the perception can be radically different if the objects of interest are rather suggested to the user by pre-computing relevant clusters of instances. Impressive results involving very small objects discovered in a web collection of 110K images are demonstrated through a simple interactive GUI.

How-to train effective classifiers on huge amount of multimedia data is clearly a major challenge that is attracting more and more research works across several communities. Less efforts however are spent on the counterpart scalability issue: how to apply big trained models efficiently on huge non annotated media collections ? In this paper, we address the problem of speeding-up the prediction phase of linear Support Vector Machines via Locality Sensitive Hashing. We propose building efficient hash based classifiers that are applied in a first stage in order to approximate the exact results and filter the hypothesis space. Experiments performed with millions of one-against-one classifiers show that the proposed hash-based classifier can be more than two orders of magnitude faster than the exact classifier with minor losses in quality.

Efficiently constructing the K-Nearest Neighbor Graph (K-NNG) of large and high dimensional datasets is crucial for many applications with feature-rich objects, such as images or other multimedia content. In this paper we investigate the use of high dimensional hashing methods for efficiently approximating the K-NNG, notably in distributed environments. We first discuss the importance of balancing issues on the performance of such approaches and show why the baseline approach using Locality Sensitive Hashing does not perform well. Our new KNN-join method is based on RMMH, a recently introduced hash function family based on randomly trained classifiers. We show that the resulting hash tables are much more balanced and that the number of resulting collisions can be greatly reduced without degrading quality. We further improve the load balancing of our distributed approach by designing a parallelized local join algorithm, implemented within the MapReduce framework.

This paper presents a visual-based media event detection system based on the automatic discovery of the most circulated images across the main news media (news websites, press agencies, TV news and newspapers). Its main originality is to rely on the transmedia contextual information to denoise the raw visual detections and consequently focus on the most salient transmedia events.

This paper presents a scalable method for automatically discovering frequent visual objects in large multimedia collections even if their size is very small. It first formally revisits the problem of mining or discovering such objects, and then generalizes two kinds of existing methods for probing candidate object seeds: weighted adaptive sampling and hashing-based methods. The idea is that the collision frequencies obtained with hashing-based methods can actually be converted into a prior probability density function given as input to a weighted adaptive sampling algorithm. This allows for an evaluation of any hashing scheme effectiveness in a more generalized way, and a comparison with other priors, e.g. guided by visual saliency concerns. We then introduce a new hashing strategy, working first at the visual level, and then at the geometric level. This strategy allows us to integrate weak geometric constraints into the hashing phase itself and not only neighborhood constraints as in previous works. Experiments conducted on a new dataset introduced in this paper will show that using this new hashing-based prior allows a drastic reduction of the number of tentative probes required to discover small objects instantiated several times in a large dataset.

Following the success of hashing methods for multidimensional indexing, more and more works are interested in embedding visual feature space in compact hash codes. Such approaches are not an alternative to using index structures but a complementary way to reduce both the memory usage and the distance computation cost. Several data dependent hash functions have notably been proposed to closely fit data distribution and provide better selectivity than usual random projections such as LSH. However, improvements occur only for relatively small hash code sizes up to 64 or 128 bits. As discussed in the paper, this is mainly due to the lack of independence between the produced hash functions. We introduce a new hash function family that attempts to solve this issue in any kernel space. Rather than boosting the collision probability of close points, our method focus on data scattering. By training purely random splits of the data, regardless the closeness of the training samples, it is indeed possible to generate consistently more independent hash functions. On the other side, the use of large margin classifiers allows to maintain good generalization performances. Experiments show that our new Random Maximum Margin Hashing scheme (RMMH) outperforms four state-of-the-art hashing methods, notably in kernel spaces.

State-of-the-art large-scale object retrieval systems usually combine efficient Bag-of-Words indexing models with a spatial verification re-ranking stage to improve query performance. In this paper we propose to directly discover spatially verified visual words as a batch process. Contrary to previous related methods based on feature sets hashing or clustering, we suggest not trading recall for efficiency by sticking on an accurate two-stage matching strategy. The problem then rather becomes a sampling issue: how to effectively and efficiently select relevant query regions while minimizing the number of tentative probes? We therefore introduce an adaptive weighted sampling scheme, starting with some prior distribution and iteratively converging to unvisited regions. Interestingly, the proposed paradigm is generalizable to any input prior distribution, including specific visual concept detectors or efficient hashing-based methods. We show in the experiments that the proposed method allows to discover highly interpretable visual words while providing excellent recall and image representativity.

Understanding the results returned by automatic visual concept detectors is often a tricky task making users uncomfortable with these technologies. In this paper we attempt to build humanly interpretable visual models, allowing the user to visually understand the underlying semantic. We therefore propose a supervised multiple instance learning algorithm that selects as few as possible discriminant local features for a given object category. The method finds its roots in the lasso theory where a L1-regularization term is introduced in order to constraint the loss function, and subsequently produce sparser solutions. Efficient resolution of the lasso path is achieved through a boosting-like procedure inspired by BLasso algorithm. Quantitatively, our method achieves similar performance as current state-of-the-art, and qualitatively, it allows users to construct their own model from the original set of patches learned, thus allowing for more compound semantic queries.

Organizing media according to real-life events is attracting interest in the multimedia community. Event-centric indexing approaches are very promising for discovering more complex relationships between data. In this paper we introduce a new visual-based method for retrieving events in photo collections, typically in the context of User Generated Contents. Given a query event record, represented by a set of photos, our method aims to retrieve other records of the same event, typically generated by distinct users. Similarly to what is done in state-of-the-art object retrieval systems, we propose a two-stage strategy combining an efficient visual indexing model with a spatiotemporal verification re-ranking stage to improve query performance. For efficiency and scalability concerns, we implemented the proposed method according to the MapReduce programming model using Multi-Probe Locality Sensitive Hashing. Experiments were conducted on LastFM-Flickr dataset for distinct scenarios, including event retrieval, automatic annotation and tags suggestion. As one result, our method is able to suggest the correct event tag over 5 suggestions with a 72% success rate.

This demo presents a crowdsourcing web application dedicated to the access of botanical knowledge through automated identification of plant species by visual content. Inspired by citizen sciences, our aim is to speed up the collection and integration of raw botanical observation data, while providing to potential users an easy and efficient access to this botanical knowledge. The result presented during the demo is an enjoying application where anyone can play to shoot fresh cut leaves and observe the relevance of species suggested in spite of various visual difficult queries.

This paper presents a new content-based retrieval framework applied to logo retrieval in large natural image collections. The first contribution is a new challenging dataset, called BelgaLogos, which was created in collaboration with professionals of a press agency, in order to evaluate logo retrieval technologies in real-world scenarios. The second and main contribution is a new visual query expansion method using an a contrario thresholding strategy in order to improve the accuracy of expanded query images. Whereas previous methods based on the same paradigm used a purely hand tuned fixed threshold, we provide a fully adaptive method enhancing both genericity and effectiveness. This new technique is evaluated on both OxfordBuilding dataset and our new BelgaLogos dataset.

This paper presents an efficient local features boosting strategy for interactive objects retrieval tasks such as on-line supervised learning or relevance feedback. The prediction time complexity of most existing methods is indeed usually linear in dataset size since the retrieval works by applying a trained classifier on the images of the dataset one by one. In our method, the trained classifier can be computed directly on the whole dataset in sublinear time thanks to distance-based weak classifiers. The idea is to speed-up drastically the prediction of each weak classifier on the whole dataset by performing approximate range queries with an efficient similarity search structure. Experiments on Caltech 256 dataset show that the technique is up to 250 times faster than the naive exhaustive method. Thanks to this efficiency improvement, we developed a relevance feedback mechanism on image regions freely selected by the user and we show how it improves the effectiveness of the retrieval.

Efficient high-dimensional similarity search structures are essential for building scalable content-based search systems on feature-rich multimedia data. In the last decade, Locality Sensitive Hashing (LSH) has been proposed as indexing technique for approximate similarity search. Among the most recent variations of LSH, multi-probe LSH techniques have been proved to overcome the overlinear space cost drawback of common LSH. Multi-probe LSH is built on the well-known LSH technique, but it intelligently probes multiple buckets that are likely to contain query results in a hash table. Our method is inspired by previous work on probabilistic similarity search structures and improves upon recent theoretical work on multi-probe and query adaptive LSH. Whereas these methods are based on likelihood criteria that a given bucket contains query results, we define a more reliable a posteriori model taking account some prior about the queries and the searched objects. This prior knowledge allows a better quality control of the search and a more accurate selection of the most probable buckets. We implemented a nearest neighbors search based on this paradigm and performed experiments on different real visual features datasets. We show that our a posteriori scheme outperforms other multi-probe LSH while offering a better quality control. Comparisons to the basic LSH technique show that our method allows consistent improvements both in space and time efficiency.

Content-based copy retrieval (CBCR) aims at retrieving in a database all the modified versions or the previous versions of a given candidate object. In this paper, we present a copy-retrieval scheme based on local features that can deal with very large databases both in terms of quality and speed. We first propose a new approximate similarity search technique in which the probabilistic selection of the feature space regions is not based on the distribution in the database but on the distribution of the features distortion. Since our CBCR framework is based on local features, the approximation can be strong and reduce drastically the amount of data to explore. Furthermore, we show how the discrimination of the global retrieval can be enhanced during its post-processing step, by considering only the geometrically consistent matches. This framework is applied to robust video copy retrieval and extensive experiments are presented to study the interactions between the approximate search and the retrieval efficiency. Largest used database contains more than 1 billion local features corresponding to 30000 h of video

Recent methods based on interest points and local fingerprints have been proposed to perform robust CBCD (content-based copy detection) of images and video. They include two steps: the search for similar local fingerprints in the database (DB) and a voting strategy that merges all the local results in order to perform a global decision. In most image or video retrieval systems, the search for similar features in the DB is performed by a geometrical query in a multidimensional index structure. Recently, the paradigm of approximate k-nearest neighbors query has shown that trading quality for time can be widely profitable in that context. In this paper, we evaluate a new approximate search paradigm, called statistical similarity search (S/sup 3/) in a complete CBCD scheme based on video local fingerprints. Experimental results show that these statistical queries allow high performance gains compared to classical e-range queries and that trading quality for time during the search does not degrade seriously the global robustness of the system, even with very large DBs including more than 20,000 hours of video.

In many image or video retrieval systems, the search for similar objects in the database includes a spatial access method to a multidimensional feature space. This step is generally considered as a problem independent of the features and the similarity type. The well known multidimensional nearest neighbor search has also been widely studied by the database community as a generic method. We propose a novel strategy dedicated to pseudo-invariant features retrieval and more specifically applied to content based copy identification. The range of a query is computed during the search according to deviation statistics between original and observed features. Furthermore, this approximate search range is directly mapped onto a Hilbert space-filling curve, allowing an efficient access to the database. Experimental results give excellent response times for very large databases both on synthetic and real data. This work is used in a TV monitoring system including more than 13000 hours of video in the reference database.