Diabetic macular edema (DME) and age-related macular degeneration (AMD) are major causes of vision impairment and blindness. While many classification applications for these diseases achieve high performance, they often overlook the crucial aspect of dataset and image quality, leading to potential erroneous predictions. This study emphasizes the importance of data quality in medical image classification, specifically for retinal imaging. We propose an Optical Coherence Tomography (OCT) image quality evaluation model using the pre-trained ARNIQA (leArning distoRtion maNifold for Image Quality Assessment) model to accurately identify retinal diseases autonomously. Our methodology includes a three-class classification system utilizing two Convolutional Neural Network (CNN) models, ResNet50 and Xception, applied to three datasets: the original dataset, a subset of high-quality images, and a subset of low-quality images. Using a Tunisian OCT dataset of 2887 images, we demonstrate the efficacy of our approach, achieving $100 \%$ accuracy with highquality images.

International audience

Human action recognition (HAR) is a critical task in the field of robotics. Traditionally, HAR methods rely on either perceptual features from RGB images or skeletal features. While RGB-based features are typically represented in 2D Euclidean space, few approaches differentiate between methods developed for RGB data and those for skeletal features, often treating both as Euclidean representations. This conventional approach, which typically leverages standard deep learning techniques, limits the descriptive power of skeletal data, which naturally exhibits a tree-like structure. In this paper, we introduce a novel framework that, for the first time, utilizes skeletal data while preserving its inherent structure to fully capture its descriptive potential. Our proposed deep neural network embeds skeletal joints into hyperbolic space, followed by a spatio-temporal processing framework that incorporates established transformations to optimize performance while maintaining the advantages of hyperbolic analysis. Extensive experiments on publicly available datasets, including UAV-Human, UAV-Gesture, and DHG 14/28, demonstrate that our approach achieves state-of-the-art results, underscoring its ability to enhance robotic systems’ performance in dynamic environments.

Diabetic macular edema (DME) and Age-Related Macular Degeneration (AMD) are two of the most common disorders that can cause blindness in a population and primarily cause retinal degradation. The application of multiple deep learning algorithms on Optical Coherence Tomography OCT) images to detect these disorders demonstrates excellent performance. However, because these algorithms include black box features, medical professionals are hesitant to fully trust the results. To address these challenges, we present a modified convolutional neural network based on the xception architecture for diagnosing DME and AMD using optical coherence tomography (OCT) images. To demonstrate the model’s transparency and trustworthiness, we used the Grad-CAM technique, which incorporates Explainable AI into the research and improves model interpretability. This technique assists medical specialists in demystifying deep learning algorithms and obtaining more information about the critical areas in OCT images used for prediction. The proposed model achieved an accuracy of 99.87%, a precision of 99.67%, and a recall of 98.29% on a dataset of 934 images.

Human action recognition with UAVs has garnered high interest due to its significant impact on various fields. This shift necessitates the creation of comprehensive and demanding benchmarks, crucial for the development and assessment of UAV-centric human behavior analysis models. However, the manifold-based approaches in the context of UAV-human action recognition face substantial limitations, given the task’s novelty and inherent complexities.This paper presents a novel approach to UAV-human action recognition, employing skeletal-based features known for their resilience in the face of these challenges. The methodology hinges on a deep neural network capable of capturing the intricate spatial and temporal facets of human actions, resulting in the creation of Semi-Positive Definite (SPD) matrix representations. These SPD representations then serve as the foundation for action classification using a classifier module.To gauge the efficacy of our approach, we conduct rigorous evaluations using the publicly available UAV-Human action recognition dataset and UAV-Gesture dataset. Our results demonstrate the state-of-the-art performance achieved by our method, highlighting its potential to advance UAV-based human action recognition significantly.

This paper addresses the challenge of utilizing learningbased algorithms for 3D scene reconstruction on resourceconstrained end-user devices. Although integrating deep learning methods into the reconstruction pipeline has demonstrated superior performance to classical techniques, the resulting large models could be impractical for resource-limited devices. We propose an efficient solution by introducing a method to compress deep learning models used in 3D reconstruction workflows. Our approach, named DRAFT, employs knowledge distillation (KD), adapted and extended for complex feature and context extraction tasks related to optical flow. New distillation components based on algebraic signpattern matrices (SPM) and inertia enhance the KD process. Empirical validation on KITTI and Sintel benchmark datasets reveals that DRAFT consistently achieves comparable or superior performance to state-of-the-art models such as RAFT, FlowID, GMFlow, and Anyflow while significantly reducing model size. This contribution enhances the feasibility of deploying learning-based 3D scene reconstruction frameworks on edge systems. It contributes to the discourse on resource-efficient deep learning methodologies, particularly for optical flow and stereo matching. Our code is available at https://github.com/christian-tchenko/DRAFT.git.

International audience

Despite the success of Siamese networks in image indexing, face recognition, and signature verification, there has been little research on designing their architectural space compared to convolutional neural networks (CNNs). This work aims to automate the design process of Siamese network architectures and improve their performance in tasks that involve image similarity computing such as indexing and retrieval. To achieve this goal, in contrast with the current literature that focuses on improving the design of the backbone CNN, we use Differentiable Neural Architecture Search (DNAS) to explore the architecture of the Multi-Layer Perceptron (MLP) component of siamese networks, namely the projector and/or predictor heads. The main objective of these MLPs is to enhance the ability of backbone CNNs to learn strong representations from unlabeled data. Using a well-known contrastive learning framework (SimCLR) as a baseline, we show that our approach managed to improve performance on several computer vision tasks such as image classification (ImageNet) and content-based image retrieval (INRIA Holidays).

Action segmentation is a challenging task that requires accurate parsing and labeling of each action. There are two types of methods for action segmentation. The first type primarily focuses on extracting high-quality features from videos, while the second type focuses on combining textual and perceptual features through multimodal fusion. However, both types of methods have their limitations. The first type is limited to a single modality and does not leverage multimodal information, while the second type, although promising, is restricted by the language used to describe the actions in the texts. To solve these problems, we propose in this paper, an Action Text Diffusion Prior Network (ATDPN) which simultaneously improves the quality of the extracted visual features (by introducing a Video-level Diffusion Prior Sampling) and integrates the textual information to fullest extent. This leads to superior action segmentation results. Our experiments performed on GTEA dataset demonstrate the effective feature extraction ability of ATDPN.

It has become mainstream in computer vision and other machine learning domains to reuse backbone networks pretrained on large datasets as preprocessors. Typically, the last layer is replaced by a shallow learning machine of sorts; the newly-added classification head and (optionally) deeper layers are fine-tuned on a new task. Due to its strong performance and simplicity, a common pre-trained backbone network is ResNet152. However, ResNet152 is relatively large and induces inference latency. In many cases, a compact and efficient backbone with similar performance would be preferable over a larger, slower one. This paper investigates techniques to reuse a pre-trained backbone with the objective of creating a smaller and faster model. Starting from a large ResNet152 backbone pre-trained on ImageNet, we first reduce it from 51 blocks to 5 blocks, reducing its number of parameters and FLOPs by more than 6 times, without significant performance degradation. Then, we split the model after 3 blocks into several branches, while preserving the same number of parameters and FLOPs, to create an ensemble of sub-networks to improve performance. Our experiments on a large benchmark of 40 image classification datasets from various domains suggest that our techniques match the performance (if not better) of “classical backbone fine-tuning” while achieving a smaller model size and faster inference speed.

Retinal pathologies like choroidal neovascularization (CNV), drusen, and diabetic macular edema (DME) can give rise to microvascular alterations in the retina, ultimately resulting in vision impairment. The manual detection of these diseases poses a significant challenge and necessitates specialized medical expertise. To address this challenge, our study introduces novel deep learning methods for the detection of these ocular pathologies automatically and based on optical coherence tomography (OCT) scans. In our experimental setup, we utilized a dataset comprising 6000 OCT images sourced from the publicly available Kaggle dataset. Through comprehensive evaluations, our study revealed that the implementation of a bilinear convolutional neural network (B-CNN) yielded the highest classification score, surpassing the accuracy achieved by alternative models. Furthermore, when compared to other deep learning networks, our proposed approach showcased superior performance in the early diagnosis of these three ocular diseases.

Abstract 3D point clouds can represent complex 3D objects of arbitrary topologies and with fine‐grained details. They are, however, hard to regress from images using convolutional neural networks, making tasks such as 3D reconstruction from monocular RGB images challenging. In fact, unlike images and volumetric grids, point clouds are unstructured and thus lack proper parameterization, which makes them difficult to process using convolutional operations. Existing point‐based 3D reconstruction methods that tried to address this problem rely on complex end‐to‐end architectures with high computational costs. Instead, we propose in this paper a novel mechanism that decouples the 3D reconstruction problem from the structure (or parameterization) learning task, making the 3D reconstruction of objects of arbitrary topologies tractable and thus easier to learn. We achieve this using a novel Teacher‐Student network where the Teacher learns to structure the point clouds. The Student then harnesses the knowledge learned by the Teacher to efficiently regress accurate 3D point clouds. We train the Teacher network using 3D ground‐truth supervision and the Student network using the Teacher's annotations. Finally, we employ a novel refinement network to overcome the upper‐bound performance that is set by the Teacher network. Our extensive experiments on ShapeNet and Pix3D benchmarks, and on in‐the‐wild images demonstrate that the proposed approach outperforms previous methods in terms of reconstruction accuracy and visual quality.

Abstract We propose a deep reinforcement learning‐based solution for the 3D reconstruction of objects of complex topologies from a single RGB image. We use a template‐based approach. However, unlike previous template‐based methods, which are limited to the reconstruction of 3D objects of fixed topology, our approach learns simultaneously the geometry and topology of the target 3D shape in the input image. To this end, we propose a neural network that learns to deform a template to fit the geometry of the target object. Our key contribution is a novel reinforcement learning framework that enables the network to also learn how to adjust, using pruning operations, the topology of the template to best fit the topology of the target object. We train the network in a supervised manner using a loss function that enforces smoothness and penalizes long edges in order to ensure high visual plausibility of the reconstructed 3D meshes. We evaluate the proposed approach on standard benchmarks such as ShapeNet, and in‐the‐wild using unseen real‐world images. We show that the proposed approach outperforms the state‐of‐the‐art in terms of the visual quality of the reconstructed 3D meshes, and also generalizes well to out‐of‐category images.

Vehicle re-identification (re-ID) aims to automatically find vehicle identity from a large number of vehicle images captured from multiple cameras. Most existing vehicle re-ID approaches rely on fully supervised learning methodologies, where large amounts of annotated training data are required, which is an expensive task. In this paper, we focus our interest on semi-supervised vehicle re-ID, where each identity has a single labeled and multiple unlabeled samples in the training. We propose a framework which gradually labels vehicle images taken from surveillance cameras. Our framework is based on a deep Convolutional Neural Network (CNN), which is progressively learned using a feature anchoring regularization process. The experiments conducted on various publicly available datasets demonstrate the efficiency of our framework in re-ID tasks. Our approach with only 20% labeled data shows interesting performance compared to the state-of-the-art supervised methods trained on fully labeled data.

In this paper, we propose a new architecture to enhance dense layers with a Taylor Series Kernelized Layer (TSKL). The proposed layer expands the underlying linear kernel of dense layers to a higher-order Taylor series kernel. This kernel is able to learn more complex patterns than the linear one and thus be more discriminative. In other words, TKSL first maps input data to a higher-dimensional Reproducing Kernel Hilbert Space (RKHS). After that, it learns a linear classifier in that RKHS which corresponds to a powerful non-linear classifier in the original feature space. The mapping features to a higher-order RKHS is performed implicitly by leveraging the kernel trick and explicitly by combining multiple kernels. The experimental results demonstrate that the proposed layer outperforms the ordinary dense layer when uses in both Multilayer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs).

Human pose estimation and action recognition are related tasks since both problems are strongly dependent on the human body representation and analysis. Nonetheless, most recent methods in the literature handle the two problems separately. In this article, we propose a multi-task framework for jointly estimating 2D or 3D human poses from monocular color images and classifying human actions from video sequences. We show that a single architecture can be used to solve both problems in an efficient way and still achieves state-of-the-art or comparable results at each task while running with a throughput of more than 100 frames per second. The proposed method benefits from high parameters sharing between the two tasks by unifying still images and video clips processing in a single pipeline, allowing the model to be trained with data from different categories simultaneously and in a seamlessly way. Additionally, we provide important insights for end-to-end training the proposed multi-task model by decoupling key prediction parts, which consistently leads to better accuracy on both tasks. The reported results on four datasets (MPII, Human3.6M, Penn Action and NTU RGB+D) demonstrate the effectiveness of our method on the targeted tasks. Our source code and trained weights are publicly available at https://github.com/dluvizon/deephar.

Fully connected layer is an essential component of Convolutional Neural Networks (CNNs), which demonstrates its efficiency in computer vision tasks. The CNN process usually starts with convolution and pooling layers that first break down the input images into features, and then analyze them independently. The result of this process feeds into a fully connected neural network structure which drives the final classification decision. In this paper, we propose a Kernelized Dense Layer (KDL) which captures higher order feature interactions instead of conventional linear relations. We apply this method to Facial Expression Recognition (FER) and evaluate its performance on RAF, FER2013 and ExpW datasets. The experimental results demonstrate the benefits of such layer and show that our model achieves competitive results with respect to the state-of-the-art approaches.

Sketch-based 3D shape retrieval has received more attention by the pattern recognition and multimedia analysis community in the last few years. Sketches are simple to draw and could be an efficient tool for abstracting 3D models. However, it is very challenging to narrow the gap between 2D sketches and 3D models because of the discrepancy between their representations. Since the set of labels used to classify 3D shapes and 2D sketches is the same, the problem of sketch-based 3D shape retrieval may be reduced to a sketch classification problem and a 3D shape classification problem. By doing so, 3D shapes having the same label as the one predicted for a 2D sketch can be considered as relevant retrieved objects for the sketch. Nevertheless, the task of 2D sketch classification is also still challenging, particularly, due to the perception subjectivity of the designers, which increases intra-class variabilities. In this paper, we tackle the problem of sketch-based 3D shape retrieval and propose the dual independent classification solution which comprises two stages: First, we independently classify both 2D sketches and 3D shapes using deep neural networks. Then, we calculate pairwise cosine similarities between the softmax vectors of the query (2D sketch) and the target (3D shape). Our method has been evaluated on two commonly used datasets, namely SHREC13 and SHREC14. The results show that our method, despite its simplicity, has comparable and sometimes better performance when compared to state-of-the-art approaches.

We tackle the image reassembly problem with wide space between the fragments, in such a way that the patterns and colors continuity is mostly unusable. The spacing emulates the erosion of which the archaeological fragments suffer. We crop-square the fragments borders to compel our algorithm to learn from the content of the fragments. We also complicate the image reassembly by removing fragments and adding pieces from other sources. We use a two-step method to obtain the reassemblies: 1) a neural network predicts the positions of the fragments despite the gaps between them; 2) a graph that leads to the best reassemblies is made from these predictions. In this paper, we notably investigate the effect of branch-cut in the graph of reassemblies. We also provide a comparison with the literature, solve complex images reassemblies, explore at length the dataset, and propose a new metric that suits its specificities.

Action recognition and human pose estimation are closely related but both problems are generally handled as distinct tasks in the literature. In this work, we propose a multitask framework for jointly 2D and 3D pose estimation from still images and human action recognition from video sequences. We show that a single architecture can be used to solve the two problems in an efficient way and still achieves state-of-the-art results. Additionally, we demonstrate that optimization from end-to-end leads to significantly higher accuracy than separated learning. The proposed architecture can be trained with data from different categories simultaneously in a seamlessly way. The reported results on four datasets (MPII, Human3.6M, Penn Action and NTU) demonstrate the effectiveness of our method on the targeted tasks.

Archaeologists are in dire need of automated object reconstruction methods. Fragments reassembly is close to puzzle problems, which may be solved by computer vision algorithms. As they are often beaten on most image related tasks by deep learning algorithms, we study a classification method that can solve jigsaw puzzles. In this paper, we focus on classifying the relative position: given a couple of fragments, we compute their local relation (e.g. on top). We propose several enhancements over the state of the art in this domain, which is outperformed by our method by 25%. We propose an original dataset composed of pictures from the Metropolitan Museum of Art. We propose a greedy reconstruction method based on the predicted relative positions.

Abstract We propose a framework for statistical modeling of the 3D geometry and topology of botanical trees. We treat botanical trees as points in a tree‐shape space equipped with a proper metric that captures the geometric and the topological differences between trees. Geodesics in the tree‐shape space correspond to the optimal sequence of deformations, i.e. bending, stretching, and topological changes, which align one tree onto another. In this way, the 3D tree modeling and synthesis problem becomes a problem of exploring the tree‐shape space either in a controlled fashion, using statistical regression, or randomly by sampling from probability distributions fitted to populations in the tree‐shape space. We show how to use this framework for (1) computing statistical summaries, e.g. the mean and modes of variations, of a population of botanical trees, (2) synthesizing random instances of botanical trees from probability distributions fitted to a population of botanical trees, and (3) modeling, interactively, 3D botanical trees using a simple sketching interface. The approach is fast and only requires as input 3D botanical tree models with a known upright orientation.

We propose an algorithm for generating novel 3D tree model variations from existing ones via geometric and structural blending. Our approach is to treat botanical trees as elements of a tree-shape space equipped with a proper metric that quantifies geometric and structural deformations. Geodesics, or shortest paths under the metric, between two points in the tree-shape space correspond to optimal deformations that align one tree onto another, including the possibility of expanding, adding, or removing branches and parts. Central to our approach is a mechanism for computing correspondences between trees that have different structures and a different number of branches. The ability to compute geodesics and their lengths enables us to compute continuous blending between botanical trees, which, in turn, facilitates statistical analysis, such as the computation of averages of tree structures. We show a variety of 3D tree models generated with our approach from 3D trees exhibiting complex geometric and structural differences. We also demonstrate the application of the framework in reflection symmetry analysis and symmetrization of botanical trees.

Abstract We introduce co‐variation analysis as a tool for modeling the way part geometries and configurations co‐vary across a family of man‐made 3D shapes. While man‐made 3D objects exhibit large geometric and structural variations, the geometry, structure, and configuration of their individual components usually do not vary independently from each other but in a correlated fashion. The size of the body of an airplane, for example, constrains the range of deformations its wings can undergo to ensure that the entire object remains a functionally‐valid airplane. These co‐variation constraints, which are often non‐linear, can be either physical, and thus they can be explicitly enumerated, or implicit to the design and style of the shape family. In this article, we propose a data‐driven approach, which takes pre‐segmented 3D shapes with known component‐wise correspondences and learns how various geometric and structural properties of their components co‐vary across the set. We demonstrate, using a variety of 3D shape families, the utility of the proposed co‐variation analysis in various applications including 3D shape repositories exploration and shape editing where the propagation of deformations is guided by the co‐variation analysis. We also show that the framework can be used for context‐guided orientation of objects in 3D scenes.

State-of-the-art 3D shape classification and retrieval algorithms, hereinafter referred to as shape analysis, are often based on comparing signatures or descriptors that capture the main geometric and topological properties of 3D objects. None of the existing descriptors, however, achieve best performance on all shape classes. In this article, we explore, for the first time, the usage of covariance matrices of descriptors, instead of the descriptors themselves, in 3D shape analysis. Unlike histogram -based techniques, covariance-based 3D shape analysis enables the fusion and encoding of different types of features and modalities into a compact representation. Covariance matrices, however, are elements of the non-linear manifold of symmetric positive definite (SPD) matrices and thus \BBL2 metrics are not suitable for their comparison and clustering. In this article, we study geodesic distances on the Riemannian manifold of SPD matrices and use them as metrics for 3D shape matching and recognition. We then: (1) introduce the concepts of bag of covariance (BoC) matrices and spatially-sensitive BoC as a generalization to the Riemannian manifold of SPD matrices of the traditional bag of features framework, and (2) generalize the standard kernel methods for supervised classification of 3D shapes to the space of covariance matrices. We evaluate the performance of the proposed BoC matrices framework and covariance -based kernel methods and demonstrate their superiority compared to their descriptor-based counterparts in various 3D shape matching, retrieval, and classification setups.

Several descriptors have been proposed in the past for 3D shape analysis, yet none of them achieves best performance on all shape classes. In this paper we propose a novel method for 3D shape analysis using the covariance matrices of the descriptors rather than the descriptors themselves. Covariance matrices enable efficient fusion of different types of features and modalities. They capture, using the same representation, not only the geometric and the spatial properties of a shape region but also the correlation of these properties within the region. Covariance matrices, however, lie on the manifold of Symmetric Positive Definite (SPD) tensors, a special type of Riemannian manifolds, which makes comparison and clustering of such matrices challenging. In this paper we study covariance matrices in their native space and make use of geodesic distances on the manifold as a dissimilarity measure. We demonstrate the performance of this metric on 3D face matching and recognition tasks. We then generalize the Bag of Features paradigm, originally designed in Euclidean spaces, to the Riemannian manifold of SPD matrices. We propose a new clustering procedure that takes into account the geometry of the Riemannian manifold. We evaluate the performance of the proposed Bag of Covariance Matrices framework on 3D shape matching and retrieval applications and demonstrate its superiority compared to descriptor-based techniques.

The need for efficient tools to index and retrieve 3D content becomes even more acute. This paper presents a fully automatic 3D-object retrieval method. It consists of two main steps namely shape signature extraction to describe the shape of objects, and similarity computing to compute similarity between objects. In the first step (signature extraction), we use a shape descriptor called geodesic cords. This descriptor can be seen as a probability distribution sampled from a shape function. In the second step (similarity computing), a global distance, based on belief function theory, is computed between each pair wise of descriptors corresponding respectively to an object query and an object from a given database. Experiments on commonly-used benchmarks demonstrate that our method obtains competitive performance compared to 3D-object retrieval methods from the state-of-the-art.

The intent of 3D-model classification is to find categories of similar objects according to their shapes. This task is a challenging and important problem in 3D-mining and shape processing. In this paper, we present a novel method to categorize 3D-objects based on view-based descriptors. The proposed method goes into two stages. The first stage corresponds to the training in which 3D-objects in the same category are processed and a set of representative 2D views is selected, The second stage corresponds to the labelling in which unknown objects are classified using a belief based classifier. The experimental results obtained on the Shrec07 datasets show that the system efficiently performs in categorizing 3D-models.

Elastic shape analysis on non-linear Riemannian manifolds provides an efficient and elegant way for simultaneous comparison and registration of non-rigid shapes. In such formulation, shapes become points on some high dimensional shape space. A geodesic between two points corresponds to the optimal deformation needed to register one shape onto another. The length of the geodesic provides a proper metric for shape comparison. However, the computation of geodesics, and therefore the metric, is computationally very expensive as it involves a search over the space of all possible rotations and re- parameterization. This problem is even more important in shape retrieval scenarios where the query shape is compared to every element in the collection to search. In this paper, we propose a new procedure for metric approximation using the framework of kernel functions. We will demonstrate that this provides a fast approximation of the metric while preserving its invariance properties.

In this paper, we present an efficient 3D object retrieval method invariant to scale, orientation and pose. Our approach is based on the dense extraction of discriminative local descriptors extracted from 2D views. We aggregate the descriptors into a single vector signature using tensor products. The similarity between 3D models can then be efficiently computed with a simple dot product. Experiments on the SHREC12 commonly-used benchmark demonstrate that our approach obtains superior performance in searching for generic shapes.

Grouping 3D objects into (semantically) meaningful categories is a challenging and important problem in 3D mining and shape processing. Here, we present a novel approach to categorize 3D objects. The method described in this article, is a belief-function-based approach and consists of two stages: the training stage, where 3D objects in the same category are processed and a set of representative parts is constructed, and the labeling stage, where unknown objects are categorized. The experimental results obtained on the Tosca-Sumner and the Shrec07 datasets show that the system efficiently performs in categorizing 3D models.

In this paper, we present a new method for 3D-shape categorization using Bag-of-Feature techniques (BoF). This method is based on vector quantization of invariant descriptors of 3D-object patches. We analyze the performance of two wellknown classifiers: the Naïve Bayes and the SVM. The results show the effectiveness of our approach and prove that the method is robust to non-rigid and deformable shapes, in which the class of transformations may be very wide due to the capability of such shapes to bend and assume different forms.

The 3D-shape matching problem plays a crucial role in many applications, such as indexing or modeling, by example. Here, we present a novel approach to matching 3D objects in the presence of nonrigid transformation and partially similar models. In this paper, we use the representation of surfaces by 3D curves extracted around feature points. Indeed, surfaces are represented with a collection of closed curves, and tools from shape analysis of curves are applied to analyze and to compare curves. The belief functions are used to define a global distance between 3D objects. The experimental results obtained on the TOSCA and the SHREC07 data sets show that the system performs efficiently in retrieving similar 3D models.

In this paper, we propose a new approach for 3D-shape matching. This approach encloses an off-line step and an on-line step. In the off-line one, an alphabet, of which any shape can be composed, is constructed. First, 3D-objects are subdivided into a set of 3D-parts. The subdivision consists to extract from each object a set of feature points with associated curves. Then the whole set of 3D-parts is clustered into different classes from a semantic point of view. After that, each class is modeled by a Hidden Markov Model (HMM). The HMM, which represents a character in the alphabet, is trained using the set of curves corresponding to the class parts. Hence, any 3D-object can be represented by a set of characters. The on-line step consists to compare the set of characters representing the 3D-object query and that of each object in the given dataset. The experimental results obtained on the TOSCA dataset show that the system efficiently performs in retrieving similar 3D-models.