Welsh is a linguistically rich yet under-resourced minority language. Despite its cultural significance, automated fluency assessment remains largely unexplored due to limited datasets and tools. Existing models focus on high-resource languages, leaving Welsh without sufficient multi-modal resources. To address this, we introduce CymruFluency, the first 4D dataset for Welsh fluency assessment, capturing both audio and 3D lip movements with expert-annotated fluency scores. Building on this, we propose a multi-modal fluency classification framework that combines audio features (mel spectrograms) and manually annotated 3D lip landmarks. Our fusion approach significantly improves fluency prediction over unimodal models, emphasizing the critical role of 3D lip dynamics in Welsh learning. This research advances minority language processing by integrating articulatory features into fluency evaluation, offering a powerful tool for Welsh language learning, assessment, and preservation. Project page: https://github.com/arvinsingh/CymruFluency .

Scale-invariance, good localization and robustness to noise and distortions are the main properties that a local feature detector should possess. Most existing local feature detectors find excessive unstable feature points that increase the number of keypoints to be matched and the computational time of the matching step. In this paper, we show that robust and accurate keypoints exist in the specific scale-space domain. To this end, we first formulate the superimposition problem into a mathematical model and then derive a closed-form solution for multiscale analysis. The model is formulated via difference-of-Gaussian (DoG) kernels in the continuous scale-space domain, and it is proved that setting the scale-space pyramid's blurring ratio and smoothness to 2 and 0.627, respectively, facilitates the detection of reliable keypoints. For the applicability of the proposed model to discrete images, we discretize it using the undecimated wavelet transform and the cubic spline function. Theoretically, the complexity of our method is less than 5% of that of the popular baseline Scale Invariant Feature Transform (SIFT). Extensive experimental results show the superiority of the proposed feature detector over the existing representative hand-crafted and learning-based techniques in accuracy and computational time. The code and supplementary materials can be found at https://github.com/mogvision/FFD.

In this paper, we bring together two divergent strands of research: photometric face capture and statistical 3D face appearance modelling. We propose a novel lightstage capture and processing pipeline for acquiring ear-to-ear, truly intrinsic diffuse and specular albedo maps that fully factor out the effects of illumination, camera and geometry. Using this pipeline, we capture a dataset of 50 scans and combine them with the only existing publicly available albedo dataset (3DRFE) of 23 scans. This allows us to build the first morphable face albedo model. We believe this is the first statistical analysis of the variability of facial specular albedo maps. This model can be used as a plug in replacement for the texture model of the Basel Face Model and we make our new albedo model publicly available. We ensure careful spectral calibration such that our model is built in a linear sRGB space, suitable for inverse rendering of images taken by typical cameras. We demonstrate our model in a state of the art analysis-by-synthesis 3DMM fitting pipeline, are the first to integrate specular map estimation and outperform the Basel Face Model in albedo reconstruction.

Video summarization is useful to find a concise representation of the original video, nevertheless its evaluation is somewhat challenging. This paper proposes a simple and efficient method for precisely evaluating the video summaries produced by the existing techniques. This method includes two steps. The first step is to establish a set of matched frames between automatic summary (AT) and the ground truth summary (GT) through two-way search, in which the similarity between two frames are measured using correlation coefficient. The second step is to estimate the consistency among these established matches, so that the difference among these frames in the AT and GT are preserved respectively. To accomplish this, a compatibility matrix is built based on the features extracted from each of these frames. The consistency values among these matched frames are estimated as the eigenvector of this matrix corresponding to the maximum eigenvalue. Such matched frames with a small enough consistency value will be rejected, leading to more accurate performance estimation of the video summarization techniques. Experimental results based on a publicly accessible dataset shows that the proposed method is effective in finding true matches and provide more realistic measurement of the performance for various techniques.

With the rise of digital content and web-based technologies, archaeologists and heritage organisations are increasingly striving to produce digital records of archaeology and heritage sites. The large numbers and geographical spread of these sites means that it would be too time-consuming for any one team to survey them. To meet this challenge, the Heritage Together project has developed a web platform through which members of the public can upload their own photographs of heritage assets to be processed into 3D models using an automated photogrammetry work flow. The web platform is part of a larger project which aims to capture, create and archive digital heritage assets in conjunction with local communities in Wales, UK, with a focus on megalithic monuments. Heritage Together is a digital community and community-built archive of heritage data, developed to inspire local communities to learn more about their heritage and to help to preserve it.

We compare skin texture classification using various 2D texture descriptors and their extensions to 3D surface orientation data. We perform a multi-resolution analysis on both the 2D and 3D data. Rotation-Invariant Local Binary Patterns, Multiple Orientations Gabor Filters and Center-Symetric Autocorrelation are used to extract 2D texture features from high resolution facial skin albedo patches. For extracting texture feature directly from the corresponding normal map patches, we propose extensions of these texture measures in both the slant/tilt and tangent spaces. We compare the results of classifying facial wrinkles and pores using the 2D-based and 3D-based texture features. We use the 3DRFE dataset which consists of high resolution 3D facial scans along with the corresponding photometric and albedo images. We notice a net improvement on classifying both wrinkle and pore using the 3D orientation based features over the 2D ones.

This paper describes an improved system for locating facial features in images using constrained local models (CLM). CLM links a set of local patch classifiers via a PCA shape model for non-rigid alignment and tracking. The convex quadratic fitting (CQF) approach to CLM approximates the patch responses with quadratic functions, allowing the parameter updates to be calculated directly. The Bayesian CLM (BCLM) further extended this approach framing it as a Bayesian inference problem. We further extend the BCLM approach to enable the use of 3D shape models. A 3D shape model is preferred on theoretical grounds and improved performance is confirmed via an empirical evaluation. The extension to 3D is developed by first introducing a full similarity transform to the (linearized) 2D CQF error function. The minimization of this error function gives a set of parameter updates that can be combined with the current estimates via a compositional approach. The adaptation of the algorithm to 3D then follows directly. The resulting algorithm is evaluated on the labeled faces in the wild (LFW) dataset and the results show improved performance over both 2D BCLM and 3D CLM.

In this paper we present a reformulation of the tensorface analysis method and produce a model that is simpler (i.e. has fewer parameters), is more compact (i.e. has tighter distributions) and is less ambiguous (i.e. no 2 sets of parameters synthesise the same data vector). This is achieved by simply subtracting the fibre (row, column, etc) mean from each fibre of the training data before performing PCA analysis. Centring of tensor data via subtraction of the whole set mean is commonly used as a preprocessing step, but the fibre-centring algorithm presented here has not been suggested previously for tensorface analysis. We show how the new formulation allows an approximate linear analysis with a considerable speed improvement over previous methods. In addition, the centring allows simpler truncation of parameter vectors leading to an even more compact model. The new method is tested on image synthesis and analysis and in a simple face recognition task, in which it out performs non-centred multilinear analysis. 1

An efficient and robust facial tracking system is presented in this paper. The system is capable of distinguishing a human face from a complex background using motion and histogram based methods. We correct for variations in illumination using a mixture of local and global illumination balance techniques. We detect and track six facial feature points i.e. pupils, nostrils and lip corners using facial feature illumination, geometric characteristics and motion information. In addition, a 3D facial feature model is employed to estimate the 3D pose of the subject's head, which improves the robustness of the tracking system. This system has the advantage of automatically detecting the facial features and recovering the features lost during the tracking process. Encouraging results have been obtained using the proposed system.

Facial feature tracking is crucial in computer vision applications. In this paper, we propose a system capable of locating a human face using skin color filtering and then detecting and tracking six facial features, i.e. pupils, nostrils and lip corners, in a real time video. A 3D facial feature model is employed to estimate the 3D pose of a person's head, which improves the robustness of the tracking system. This system has the advantage of automatically detecting the facial features and recovering the features lost during the tracking process. Encouraging results have been obtained using the proposed system.

Animated talking faces can be generated from a set of predefined face and mouth shapes (visemes) by either concatenation or morphing. Each facial image corresponds to one or more phonemes, which are generated in synchrony with the visual changes. Existing implementations require a full set of facial visemes to be captured or created by an artist before the images can be animated. In this work we generate new, photo-realistic visemes from a single neutral face image by transformation using a set of prototype visemes. This allows us to generate visual speech from photographs and portraits where a full set of visemes is not available.