Online task-free continual learning requires models to learn from a non-stationary data stream without task boundaries, replay buffers, or multiple passes, while preserving previously acquired knowledge. Existing exemplar-free methods either rely on unstable geometric assumptions or synthetic feature generation, which may degrade under highly non-stationary streams. To address these limitations, we propose GCHP, a Geometry-Controlled Convex Hull Prototype Framework with adaptive structure control for online task-free continual learning. GCHP represents each class by maintaining a convex hull in a low-dimensional control space, together with its corresponding semantic prototypes in the original feature space. Upon receiving new samples, the hull is updated via expansion and contraction in the control space, and regulated prototypes are embedded into the original feature space for inference. This geometry-controlled mechanism enables stable boundary refinement under bounded prototype capacity without storing real exemplars. Extensive experiments on CIFAR-10, CIFAR-100, CORe-50, and CUB-200 show that GCHP achieves strong performance across benchmarks, despite operating in a strictly single-pass setting without using memory buffers. GCHP achieves competitive or superior performance compared to prior approaches across multiple benchmarks, outperforming existing methods on several datasets while remaining competitive on more challenging fine-grained scenarios such as CUB-200. Ablation studies further demonstrate the importance of the geometric representation and the robustness of the prototype capacity. These results underline the effectiveness of geometric consolidation for continual learning and highlight GCHP as a simple, stable, and scalable alternative for online exemplar-free scenarios. The code implementation of the proposed GCHP framework is available at https://github.com/tutc/GCHP .

Collecting annotations for dense scene object detection (OD) is notoriously labor-intensive, as real-world images often contain hundreds of small, overlapping objects. In many practical scenarios, even annotating a few images is tedious and frustrating, motivating the need for unsupervised solutions. Despite its importance, this problem remains largely unexplored, with no existing methods specifically designed to address it. Therefore, in this work, we tackle the problem of training object detectors for dense scenes without any human annotations. Leveraging GroundingDINO [1], a recent Vision-Language model (VLM), which supports open-vocabulary detection conditioned by textual prompts, we explore its potential for generating pseudo-labels (PLs) directly from images. However, standard inference of this model often fail to capture all relevant objects in crowded scenes. To overcome this problem, we introduce M ulti- S cale T iling I nference (MSTI), a strategy that applies the VLM over overlapping image tiles at multiple tile scales to enhance performance through improved coverage. We further propose a pre-filtering mechanism that adaptively determines when and where in an image, MSTI should be applied. This results in a final pipeline, AdaMSTI, that balances accuracy with computational efficiency. The results from experiments conducted on two datasets, CrowdHuman [2] and SKU110K [3], highlight the efficiency of our method: the PLs generated by the proposed method are significantly more accurate than the naive baselines with GroundingDINO [1] , leading to notable improvements in detectors trained on these labels, compared to those that were trained with Cut-and-LEaRn (CutLER) [4], the state-of-the-art unsupervised OD for curated data. Furthermore, on CrowdHuman [2] where the object counts per image vary, the full AdaMSTI pipeline demonstrates that it produces high-quality PLs while reducing unnecessary computation. Detectors trained on AdaMSTI PLs achieve higher validation performance than those trained on naive or full MSTI PLs.

In this article, we introduce an approach called coupled filters decomposition, which builds on the key observation that redundancy exists among filters in a convolutional layer, meaning that similar filters can produce partially overlapping outputs. Leveraging this insight, we propose a joint decomposition of filters using coupled tensor decompositions, specifically coupled canonical polyadic decomposition (CPD), which enables the sharing of a common factor matrix across similar filters. This joint factorization not only reduces the number of parameters but also lowers computational complexity by eliminating redundant computations. To further improve efficiency, we first cluster the filters before decomposition. The grouping relies on a custom metric based on the subspace spanned by the shared-mode factor. Within each group, the coupling constraint is less restrictive. Extensive experiments across various architectures, datasets, and tasks validate the effectiveness of our method, demonstrating its competitive performance compared to state-of-the-art model compression techniques. Our code is available for research purposes at https://codec-ai.github.io/.

In this paper, we present SLIMING (Singular vaLues-drIven autoMated filter prunING), an automated filter pruning method that uses singular values to formalize the pruning process as an optimization problem over filter tensors. Recognizing that this original formulation poses a combinatorial challenge, we propose to replace it with a two-step process that consistently uses singular values in each phase: (i) determining the pruning configuration, which specifies the number of filters to retain in each layer, and (ii) selecting the filters themselves. We show that this approach ensures the preservation of the filters' multidimensional structure throughout the pruning process. For each of these steps, we propose a straightforward algorithm to solve them. To validate each part of our approach, we performed a numerical simulation on an overparameterized synthetic toy example. Additionally, we conducted extensive simulations across eight architectures, four benchmark datasets, and four vision tasks, validating the efficacy of our framework. Our code is available for research purposes at sliming-ai.github.io.

Network compression techniques that combine tensor decompositions and pruning have shown promise in leveraging the advantages of both strategies. In this work, we propose enhanced Network cOmpRession through TensOr decompositions and pruNing (NORTON), a novel method for network compression. NORTON introduces the concept of filter decomposition, enabling a more detailed decomposition of the network while preserving the weight's multidimensional properties. Our method incorporates a novel structured pruning approach, effectively integrating the decomposed model. Through extensive experiments on various architectures, benchmark datasets, and representative vision tasks, we demonstrate the usefulness of our method. NORTON achieves superior results compared to state-of-the-art (SOTA) techniques in terms of complexity and accuracy. Our code is also available for research purposes.

A novel method for detecting rotational symmetry is addressed in this paper by introducing a new concept of semi-shapes to overcome the main problem of projection-based approaches for studying rotational symmetric properties of an arbitrary shape. It is due to the fact that in the classical approaches, projection cues are periodical with a period of π preventing exploitation of rotational properties. We then propose the profile of semi-shapes as a signature of the shape together with a simple yet efficient technique to determine the rotation symmetry of an arbitrary shape by considering the correlation of this signature and its circular shift. A new measure is also introduced to determine how good the rotational symmetry would be. Experiments on single/compound-contour shapes have clearly corroborated the efficacy of our proposal.

Abstract We present a novel framework for efficient and robust facial feature representation based upon Local Binary Pattern (LBP), called Weighted Statistical Binary Pattern, wherein the descriptors utilize the straight-line topology along with different directions. The input image is initially divided into mean and variance moments. A new variance moment, which contains distinctive facial features, is prepared by extracting rootk-th. Then, when Sign and Magnitude components along four different directions using the mean moment are constructed, a weighting approach according to the new variance is applied to each component. Finally, the weighted histograms of Sign and Magnitude components are concatenated to build a novel histogram of Complementary LBP along with different directions. A comprehensive evaluation using six public face datasets suggests that the present framework outperforms the state-of-the-art methods and achieves 98.51% for ORL, 98.72% for YALE, 98.83% for Caltech, 99.52% for AR, 94.78% for FERET, and 99.07% for KDEF in terms of accuracy, respectively. The influence of color spaces and the issue of degraded images are also analyzed with our descriptors. Such a result with theoretical underpinning confirms that our descriptors are robust against noise, illumination variation, diverse facial expressions, and head poses.

We present two novel shape signature-based reflection symmetry detection methods with their theoretical underpinning and empirical evaluation. LIP-signature and R-signature share similar beneficial properties allowing to detect reflection symmetry directions in a high-performing manner. For the shape signature of a given shape, its merit profile is constructed to detect candidates of symmetry direction. A verification process is utilized to eliminate the false candidates by addressing Radon projections. The proposed methods can effectively deal with compound shapes which are challenging for traditional contour-based methods. To quantify the symmetric efficiency, a new symmetry measure is proposed over the range [0, 1]. Furthermore, we introduce two symmetry shape datasets with a new evaluation protocol and a lost measure for evaluating symmetry detectors. Experimental results using standard and new datasets suggest that the proposed methods prominently perform compared to state of the art.

Understanding dynamic textures (DTs) is a challenge in various computer vision applications due to the negative impacts of noise, changes of environment, illumination, and scales on capturing turbulent characteristics. In this work, we propose an efficient shallow framework for DT representation by addressing the following novel concepts. First, it is the first time in DT analysis that 2D/3D Gaussian-gradient filterings are taken into account as a pre-processing step to point out robust components against those influences in effect. Second, high-order partial derivatives of the Gaussian kernels and their informative magnitudes are exploited to forcefully capture multi-order Gaussian-gradient features. Third, these gradient kernels are investigated in multi-scale analysis of different orders and standard deviations in order to enrich more useful scale-gradient information. Finally, the obtained complementary components are shallowly encoded using a simple local operator to construct robust descriptors of High-order 2D/3D Gaussian-gradient-based Features ($\mathrm{HoGF}^{2D/3D}$) against the well-known issues of DT description. Experiments for DT classification on various benchmarks have validated the interest of our approach since its performance is comparable to state-of-the-art results, including that of deep-learning methods, while it only has a small dimension.

Representation of dynamic textures (DTs), well‐known as a sequence of moving textures, is a challenging problem in video analysis due to the disorientation of motion features. Analysing DTs to make them ‘understandable’ plays an important role in different applications of computer vision. In this study, an efficient approach for DT description is proposed by addressing the following novel concepts. First, the beneficial properties of dense trajectories are exploited for the first time to efficiently describe DTs instead of the whole video. Second, two substantial extensions of local vector pattern operator are introduced to form a completed model which is based on complemented components to enhance its performance in encoding directional features of motion points in a trajectory. Finally, the authors present a new framework, called directional dense trajectory patterns, which takes advantage of directional beams of dense trajectories along with spatio‐temporal features of their motion points in order to construct dense‐trajectory‐based descriptors with more robustness. Evaluations of DT recognition on different benchmark datasets (i.e. UCLA, DynTex, and DynTex++) have verified the interest of the authors’ proposal.

A novel method for reflection symmetry detection is addressed using a projection-based approach that allows to deal effectively with additional noise, non-linear deformations, and composed shapes that are not evident for classic contour-based approaches. A new symmetry measure is also proposed to measure how good the detected symmetry is. Experiments validate the interest of our proposed method.

An efficient framework for dynamic texture (DT) representation is proposed by exploiting local features based on Local Binary Patterns (LBP) from filtered images. First, Gaussian smoothing filter is used to deal with near uniform regions and noise which are typical restrictions of LBP operator. Second, the receptive field of Difference of Gaussians (DoG), which is exploited in DT description for the first time, allows to make the descriptor more robust against the changes of environment, illumination, and scale which are main challenges in DT representation. Experimental results of DT recognition on different benchmark datasets (i.e., UCLA, DynTex, and DynTex++), which give outstanding performance compared to the state of the art, verify the interest of our proposal.

In this paper, we propose a new local descriptor for action recognition in depth images. Our proposed descriptor jointly encodes the shape and motion cues using surface normals in 4D space of depth, time, spatial coordinates and higher-order partial derivatives of depth values along spatial coordinates. In a traditional Bag-of-words (BoW) approach, local descriptors extracted from a depth sequence are encoded to form a global representation of the sequence. In our approach, local descriptors are encoded using Sparse Coding (SC) and Fisher Vector (FV), which have been recently proven effective for action recognition. Action recognition is then simply performed using a linear SVM classifier. Our proposed action descriptor is evaluated on two public benchmark datasets, MSRAction3D and MSRGesture3D. The experimental result shows the effectiveness of the proposed method on both the datasets.

In this paper, we propose a new local descriptor for action recognition in depth images. The proposed descriptor relies on surface normals in 4D space of depth, time, spatial coordinates and higher-order partial derivatives of depth values along spatial coordinates. In order to classify actions, we follow the traditional Bag-of-words (BoW) approach, and propose two encoding methods termed Multi-Scale Fisher Vector (MSFV) and Temporal Sparse Coding based Fisher Vector Coding (TSCFVC) to form global representations of depth sequences. The high-dimensional action descriptors resulted from the two encoding methods are fed to a linear SVM for efficient action classification. Our proposed methods are evaluated on two public benchmark datasets, MSRAction3D and MSRGesture3D. The experimental result shows the effectiveness of the proposed methods on both the datasets.

Measuring the degree to which a shape resembles a polygon (referred to as polygonality) is a difficult problem due to the intrinsic diversity in the form and distortion of shapes caused by digitization and similarity transformation. This paper proposes a generic approach for this problem by performing the measurement in the projection space where the Radon image of some primitive shapes, which compose the shape, becomes apparent. The obtained measures take value in (0, 1] with 1 corresponding to analytical polygons. They are robust to additive noise and boundary distortion, and is invariant to similarity transformation. The new framework generalizes existing polygonal measures such as triangularity and quadrangularity. In addition, the ability to estimate polygon's geometric quantities in the projection space allows approximating a shape by analytical polygons. The efficiency of the proposed approach is demonstrated through a number of experiments on both synthetic and real data sets.

A new action model is proposed, by revisiting local binary patterns (LBP) for dynamic texture models, applied on trajectory beams calculated on the video. The use of semi-dense trajectory field allows to dramatically reduce the computation support to essential motion information, while maintaining a large amount of data to ensure robustness of statistical bag of features action models. A new binary pattern, called Spatial Motion Pattern (SMP) is proposed, which captures self-similarity of velocity around each tracked point (particle), along its trajectory. This operator highlights the geometric shape of rigid parts of moving objects in a video sequence. SMPs are combined with basic velocity information to form the local action primitives. Then, a global representation of a space × time video block is provided by using hierarchical blockwise histograms, which allows to efficiently represent the action as a whole, while preserving a certain level of spatiotemporal relation between the action primitives. Inheriting from the efficiency and the invariance properties of both the semi-dense tracker Video extruder and the LBP-based representations, the method is designed for the fast computation of action descriptors in unconstrained videos. For improving both robustness and computation time in the case of high definition video, we also present an enhanced version of the semi-dense tracker based on the so-called super particles, which reduces the number of trajectories while improving their length, reliability and spatial distribution.

A new spatio temporal descriptor is proposed for action recognition. The action is modelled from a beam of trajectories obtained using semi dense point tracking on the video sequence. We detect the dominant points of these trajectories as points of local extremum curvature and extract their corresponding feature vectors, to form a dictionary of atomic action elements. The high density of these informative and invariant elements allows effective statistical action description. Then, human action recognition is performed using a bag of feature model with SVM classifier. Experimentations show promising results on several well-known datasets.

A multi-scale approach is proposed for polygonal representation of a digital curve by using the notion of blurred segment and a split-and-merge strategy. Its main idea is to decompose the curve into meaningful parts that are represented by detected dominant points at the appropriate scale. The method uses no threshold and can automatically decompose the curve into meaningful parts.

We propose a new circularity measure inspired from Arkin, Latecki tools of shape matching that is constructed in a tangent space. We then introduce a linear algorithm that uses this measure for circularity measuring. This method can also be regarded as a method for circular object recognition. Experimental results show the robustness of this simple method.

A new and fast method for dominant point detection and polygonal representation of a discrete curve is proposed. Starting from results of discrete geometry, the notion of maximal blurred segment of width v has been proposed, well adapted to possibly noisy and/or not connected curves. For a given width, the dominant points of a curve C are deduced from the sequence of maximal blurred segments of C in O(n log2n) time. Comparisons with other methods of the literature prove the efficacity of our approach.