Abstract Recent advances in deep learning have greatly enhanced the accuracy and scalability of animal re-identification by automating the extraction of subtle distinguishing features from images and videos. This enables large-scale, non-invasive monitoring of animal populations. This article proposes a segmentation pipeline and a re-identification model to identify animals without ground-truth IDs. The segmentation pipeline isolates animals from the background using bounding boxes and leverages the DINOv2 and Segment Anything Model 2 (SAM2) foundation models. For re-identification, Recurrence over Video Frames (RoVF) is introduced, a novel approach that employs a recurrent component based on the Perceiver transformer atop a DINOv2 image model, iteratively refining embeddings from video frames. The proposed methods are evaluated on video datasets of meerkats and polar bears (PolarBearVidID). The proposed segmentation model achieved high accuracy (94.36% and 97.26%) and IoU (73.14% and 92.77%) for meerkats and polar bears, respectively. RoVF outperformed frame- and video-based re-identification baselines, achieving a top-1 accuracy of 46.5% and 55% on masked test sets for meerkats and polar bears, respectively, as well as higher top-3 accuracy. These results highlight the potential of the proposed approach to reduce annotation burdens in future individual-based ecological studies. The code is available at https://github.com/Strong-AI-Lab/RoVF-Meerkat-Reidentification .

Computer Vision (CV) labeling problems play a pivotal role in low-level vision. For decades, it has been known that these problems can be elegantly formulated as discrete energy-minimization problems derived from probabilistic graphical models such as Markov Random Fields (MRFs). Despite recent advances in MRF inference algorithms (such as graph-cut and message-passing methods), the resulting energy-minimization problems are generally viewed as intractable. The emergence of quantum computations, which offer the potential for faster solutions to certain problems than classical methods, has led to an increased interest in utilizing quantum properties to overcome intractable problems. Recently, there has also been a growing interest in Quantum Computer Vision (QCV), hoping to provide a credible alternative/assistant to deep learning solutions. This study investigates a new Quantum Annealing-based inference algorithm for CV discrete energy minimization problems. Our contribution is focused on Stereo Matching as a significant CV labeling problem. As a proof of concept, we also use a hybrid quantum–classical solver provided by D-Wave System to compare our results with the best classical inference algorithms in the literature. Our results show that Quantum Annealing can yield promising results for Stereo Matching problems, with improved accuracy on certain stereo images and competitive performance on others.

Stereo matching techniques aim at reconstructing disparity maps from a pair of images. The use of stereo matching techniques in embedded systems is very challenging due to the complexity of the state-of-the-art algorithms. Local stereo matching algorithms are efficiently implemented on GPU and DSP. This paper presents the optimization of the One Dimension Belief Propagation (BP-1D) algorithm. BP-1D is faster than previous algorithms on monocore DSP and its implementation onto multicore DSPs is straightforward. BP-1D implemented on multicore embedded platforms out-performs previous stereo matching implementations reaching real-time performances for resolutions up to 1080p with a 10 Watts power consumption.

We propose a novel approach to justify and guide regularisation of an ill-posed one-dimensional global optimisation with multiple solutions using a massively parallel (P system) model of the solution space. Classical optimisation assumes a well-posed problem with a stable unique solution. Most of important practical problems are ill posed due to an unstable or non-unique global optimum and are regularised to get a unique best-suited solution. Whilst regularisation theory exists largely for unstable unique solutions, its recommendations are often routinely applied to inverse optical problems with essentially non-unique solutions, e.g. computer stereo vision or image segmentation, typically formulated in terms of global energy minimisation. In these cases the recommended regularisation becomes purely heuristic and does not guarantee a unique solution. As a result, classical optimisation algorithms: dynamic programming (DP) and belief propagation (BP) – meet with difficulties. Our recent concurrent propagation (CP), leaning upon the P systems paradigm, extends DP and BP to always detect whether the problem is ill posed or not and store in the ill-posed case an entire space of solutions that yield the same global optimum. This suggests a radically new path to proper regularisation: select the best-suited unique solution by exploring statistical and structural features of this space. We propose a P systems based implementation of CP and set out as a case study an application of CP to the image matching problem in stereo vision.

An increasing number of applications require 3D content. However, its creation from real-world data either necessitates expensive equipment, artistic skills, or is constrained, for example, by the range of the utilized sensors. Image-based modeling is rapidly increasing in popularity since cameras are very affordable, widely available, and have a wide image acquisition range suitable for objects of vastly different size. The technique is especially suitable for mobile robotics involving low cost equipment and robots with a light payload, for example, small UAVs. In this paper we describe a novel image-based modeling system, which produces high-quality 3D content automatically from a collection of unconstrained and uncalibrated 2D images. The system estimates camera parameters and a 3D scene geometry using Structure-from-Motion (SfM) and Bundle Adjustment techniques. The point cloud density of 3D scene components is enhanced by exploiting silhouette information of the scene. This hybrid approach dramatically improves the reconstruction of objects with few visual features, for example, unicolored objects, and improves surface smoothness. A high quality texture is created by parameterizing the reconstructed objects using a segmentation and charting approach which also works for objects which are not homeomorphic to a sphere. The resulting parameter space contains one chart for each surface segment. A texture map is created by back projecting the best fitting input images onto each surface segment, and smoothly fusing them together over the corresponding chart by using graph-cut techniques.

We compare three Computer Vision approaches to 3-D reconstruction, namely passive Binocular Stereo and active Structured Lighting and Photometric Stereo, in application to human face reconstruction for modelling virtual humans. An integrated lab environment was set up to simultaneously acquire images for 3-D reconstruction and corresponding data from a 3-D scanner. This allowed us to quantitatively compare reconstruction results to accurate ground truth. Our goal was to determine whether any current Computer Vision approach is accurate enough for practically useful 3-D facial surface reconstruction. Comparative experiments show the combination of Structured Lighting with Symmetric Dynamic Programming based Binocular Stereo has good prospects due to reasonable processing time and sufficient accuracy.

The reconstruction of 3D face models is mostly achieved by using 2D images. We compare the strengths and weaknesses of different image processing techniques for 3D face generation. It is anticipated that the optimal solution will be applied in the future for 3D face analysis and synthesis. As approaches to 3D face modelling, the paper presents: binocular stereo, using a stereo correspondence algorithm or manual triangulation; orthogonal views; photometric stereo. Photometric stereo and orthogonal views seem to provide the best rendering while keeping a reasonable time efficiency, implementation difficulty and cumbersomeness. However, 3D face acquisition techniques have not closed the gap between accuracy and cumbersomeness.

An unsupervised framework for face analysis aiming at lip tracking is presented in this paper. A colour video sequence of a speaker's face is simply acquired by a desktop camera under natural lighting conditions and without any particular make-up. After a logarithmic colour transform, a statistical segmentation process regularizes motion and hue information within a spatio-temporal neighbourhood. The hierarchical segmentation labels the different areas of the face. Results are then used to define a region of interest for each feature in the face, particularly the lip contours. Lip corners and associated characteristic points are extracted to initialise an active contours stage. Finally, a speaker's lip shape with inner and outer borders is tracked without user tuning: This unsupervised framework provides geometrical features of the face when no specific model of the speaker face is assumed.

Active contours or snakes are widely used in object segmentation for their ability to integrate feature extraction and pixel candidate linking in a single energy minimizing process. But the sensitivity to parameters values and initialization is also a widely known problem. The performance of snakes can be enhanced by better initialization close to the desired solution. We present a fine mouth region of interest (ROI) extraction using gray level image and corresponding gradient information. We link this technique with an original snake method. The automatic snakes use spatially varying coefficients to remain along its evolution in a mouth-like shape. Our experimentations on a large image database prove its robustness regarding speakers change of the ROI mouth extraction and automatic snakes algorithms. The main application of our algorithms is video-conferencing.