William A. P. Smith

Finlay G. C. Hudson, William A. P. Smith

In this paper, we introduce a novel method for zero-shot, cross-domain image retrieval. Our key contribution is a test-time Iterative Cluster-free Re-ranking process that leverages gallery-gallery feature information to establish semantic links between query and gallery images. This enables the retrieval of relevant images even when they do not exhibit similar visual features but share underlying semantic concepts. This can be combined with any pre-existing cross-domain feature extraction backbone to improve retrieval performance. However, when combined with a carefully chosen Vision Transformer backbone and combination of zero-shot retrieval losses, our approach yields state-of-the-art results on the Sketchy, TU-Berlin and QuickDraw sketch-based retrieval benchmarks. We show that our re-ranking also improves performance with other backbones and outperforms other re-ranking methods applied with our backbone. Importantly, unlike many previous methods, none of the components in our approach are engineered specifically towards the sketch-based image retrieval task - it can be generally applied to any cross-domain, zero-shot retrieval task. We therefore also present new results on zero-shot cartoon-to-photo and art-to-product retrieval using the Office-Home dataset. Project page: finlay-hudson.github.io/icfrr, code available at: github.com/finlay-hudson/ICFRR

James A. D. Gardner, Bernhard Egger, William A. P. Smith

Inverse rendering is an ill-posed problem. Previous work has sought to resolve this by focussing on priors for object or scene shape or appearance. In this work, we instead focus on a prior for natural illuminations. Current methods rely on spherical harmonic lighting or other generic representations and, at best, a simplistic prior on the parameters. We propose a conditional neural field representation based on a variational auto-decoder with a SIREN network and, extending Vector Neurons, build equivariance directly into the network. Using this, we develop a rotation-equivariant, high dynamic range (HDR) neural illumination model that is compact and able to express complex, high-frequency features of natural environment maps. Training our model on a curated dataset of 1.6K HDR environment maps of natural scenes, we compare it against traditional representations, demonstrate its applicability for an inverse rendering task and show environment map completion from partial observations. A PyTorch implementation, our dataset and trained models can be found at jadgardner.github.io/RENI.

James A. D. Gardner, Evgenii Kashin, Bernhard Egger et al.

Inverse rendering of outdoor scenes from unconstrained image collections is a challenging task, particularly illumination/albedo ambiguities and occlusion of the illumination environment (shadowing) caused by geometry. However, there are many cues in an image that can aid in the disentanglement of geometry, albedo and shadows. Whilst sky is frequently masked out in state-of-the-art methods, we exploit the fact that any sky pixel provides a direct observation of distant lighting in the corresponding direction and, via a neural illumination prior, a statistical cue to derive the remaining illumination environment. The incorporation of our illumination prior is enabled by a novel `outside-in' method for computing differentiable sky visibility based on a neural directional distance function. This is highly efficient and can be trained in parallel with the neural scene representation, allowing gradients from appearance loss to flow from shadows to influence the estimation of illumination and geometry. Our method estimates high-quality albedo, geometry, illumination and sky visibility, achieving state-of-the-art results on the NeRF-OSR relighting benchmark. Our code and models can be found at https://github.com/JADGardner/neusky

James A. D. Gardner, Bernhard Egger, William A. P. Smith

Inverse rendering is an ill-posed problem. Previous work has sought to resolve this by focussing on priors for object or scene shape or appearance. In this work, we instead focus on a prior for natural illuminations. Current methods rely on spherical harmonic lighting or other generic representations and, at best, a simplistic prior on the parameters. This results in limitations for the inverse setting in terms of the expressivity of the illumination conditions, especially when taking specular reflections into account. We propose a conditional neural field representation based on a variational auto-decoder and a transformer decoder. We extend Vector Neurons to build equivariance directly into our architecture, and leveraging insights from depth estimation through a scale-invariant loss function, we enable the accurate representation of High Dynamic Range (HDR) images. The result is a compact, rotation-equivariant HDR neural illumination model capable of capturing complex, high-frequency features in natural environment maps. Training our model on a curated dataset of 1.6K HDR environment maps of natural scenes, we compare it against traditional representations, demonstrate its applicability for an inverse rendering task and show environment map completion from partial observations. We share our PyTorch implementation, dataset and trained models at https://github.com/JADGardner/ns_reni

Meghna Asthana, William A. P. Smith, Patrik Huber

We propose to tackle the multiview photometric stereo problem using an extension of Neural Radiance Fields (NeRFs), conditioned on light source direction. The geometric part of our neural representation predicts surface normal direction, allowing us to reason about local surface reflectance. The appearance part of our neural representation is decomposed into a neural bidirectional reflectance function (BRDF), learnt as part of the fitting process, and a shadow prediction network (conditioned on light source direction) allowing us to model the apparent BRDF. This balance of learnt components with inductive biases based on physical image formation models allows us to extrapolate far from the light source and viewer directions observed during training. We demonstrate our approach on a multiview photometric stereo benchmark and show that competitive performance can be obtained with the neural density representation of a NeRF.

Pourya Shamsolmoali, Masoumeh Zareapoor, Eric Granger et al.

In continual learning, the primary challenge is to learn new information without forgetting old knowledge. A common solution addresses this trade-off through regularization, penalizing changes to parameters critical for previous tasks. In most cases, this regularization term is directly added to the training loss and optimized with standard gradient descent, which blends learning and retention signals into a single update and does not explicitly separate essential parameters from redundant ones. As task sequences grow, this coupling can over-constrain the model, limiting forward transfer and leading to inefficient use of capacity. We propose a different approach that separates task learning from stability enforcement via operator splitting. The learning step focuses on minimizing the current task loss, while a proximal stability step applies a sparse regularizer to prune unnecessary parameters and preserve task-relevant ones. This turns the stability-plasticity into a negotiated update between two complementary operators, rather than a conflicting gradient. We provide theoretical justification for the splitting method on the continual-learning objective, and demonstrate that our proposed solver achieves state-of-the-art results on standard benchmarks, improving both stability and adaptability without the need for replay buffers, Bayesian sampling, or meta-learning components.

Mingrui Li, William A. P. Smith, Patrik Huber

Embedding a face image to a descriptor vector using a deep CNN is a widely used technique in face recognition. Via several possible training strategies, such embeddings are supposed to capture only identity information. Information about the environment (such as background and lighting) or changeable aspects of the face (such as pose, expression, presence of glasses, hat etc.) should be discarded since they are not useful for recognition. In this paper, we present a surprising result that this is not the case. We show that non-ID attributes, as well as landmark positions and the image histogram can be recovered from the ID embedding of state-of-the-art face embedding networks (VGGFace2 and ArcFace). In fact, these non-ID attributes can be predicted from ID embeddings with similar accuracy to a prediction from the original image. Going further, we present an optimisation strategy that uses a generative model (specifically StyleGAN2 for faces) to recover images from an ID embedding. We show photorealistic inversion from ID embedding to face image in which not only is the ID realistically reconstructed but the pose, lighting and background/apparel to some extent as well.

Paul Walker, James A. D. Gardner, Andreea Ardelean et al.

Inverse rendering is an ill-posed problem, but priors like illumination priors, can simplify it. Existing work either disregards the spherical and rotation-equivariant nature of illumination environments or does not provide a well-behaved latent space. We propose a rotation-equivariant variational autoencoder that models natural illumination on the sphere without relying on 2D projections. To preserve the SO(2)-equivariance of environment maps, we use a novel Vector Neuron Vision Transformer (VN-ViT) as encoder and a rotation-equivariant conditional neural field as decoder. In the encoder, we reduce the equivariance from SO(3) to SO(2) using a novel SO(2)-equivariant fully connected layer, an extension of Vector Neurons. We show that our SO(2)-equivariant fully connected layer outperforms standard Vector Neurons when used in our SO(2)-equivariant model. Compared to previous methods, our variational autoencoder enables smoother interpolation in latent space and offers a more well-behaved latent space.

Yash Deo, Yan Jia, Toni Lassila et al.

Evaluating generative models for synthetic medical imaging is crucial yet challenging, especially given the high standards of fidelity, anatomical accuracy, and safety required for clinical applications. Standard evaluation of generated images often relies on no-reference image quality metrics when ground truth images are unavailable, but their reliability in this complex domain is not well established. This study comprehensively assesses commonly used no-reference image quality metrics using brain MRI data, including tumour and vascular images, providing a representative exemplar for the field. We systematically evaluate metric sensitivity to a range of challenges, including noise, distribution shifts, and, critically, localised morphological alterations designed to mimic clinically relevant inaccuracies. We then compare these metric scores against model performance on a relevant downstream segmentation task, analysing results across both controlled image perturbations and outputs from different generative model architectures. Our findings reveal significant limitations: many widely-used no-reference image quality metrics correlate poorly with downstream task suitability and exhibit a profound insensitivity to localised anatomical details crucial for clinical validity. Furthermore, these metrics can yield misleading scores regarding distribution shifts, e.g. data memorisation. This reveals the risk of misjudging model readiness, potentially leading to the deployment of flawed tools that could compromise patient safety. We conclude that ensuring generative models are truly fit for clinical purpose requires a multifaceted validation framework, integrating performance on relevant downstream tasks with the cautious interpretation of carefully selected no-reference image quality metrics.

Finlay G. C. Hudson, James A. D. Gardner, William A. P. Smith

Humans excel at constructing panoramic mental models of their surroundings, maintaining object permanence and inferring scene structure beyond visible regions. In contrast, current artificial vision systems struggle with persistent, panoramic understanding, often processing scenes egocentrically on a frame-by-frame basis. This limitation is pronounced in the Track Any Point (TAP) task, where existing methods fail to track 2D points outside the field of view. To address this, we introduce TAPVid-360, a novel task that requires predicting the 3D direction to queried scene points across a video sequence, even when far outside the narrow field of view of the observed video. This task fosters learning allocentric scene representations without needing dynamic 4D ground truth scene models for training. Instead, we exploit 360 videos as a source of supervision, resampling them into narrow field-of-view perspectives while computing ground truth directions by tracking points across the full panorama using a 2D pipeline. We introduce a new dataset and benchmark, TAPVid360-10k comprising 10k perspective videos with ground truth directional point tracking. Our baseline adapts CoTracker v3 to predict per-point rotations for direction updates, outperforming existing TAP and TAPVid 3D methods. Project page: https://finlay-hudson.github.io/tapvid360

James A. D. Gardner, Will Rowan, William A. P. Smith

In this paper, we introduce NeuRaLaTeX, which we believe to be the first deep learning library written entirely in LaTeX. As part of your LaTeX document you can specify the architecture of a neural network and its loss functions, define how to generate or load training data, and specify training hyperparameters and experiments. When the document is compiled, the LaTeX compiler will generate or load training data, train the network, run experiments, and generate figures. This paper generates a random 100 point spiral dataset, trains a two layer MLP on it, evaluates on a different random spiral dataset, produces plots and tables of results. The paper took 48 hours to compile and the entire source code for NeuRaLaTeX is contained within the source code of the paper. We propose two new metrics: the Written In Latex (WIL) metric measures the proportion of a machine learning library that is written in pure LaTeX, while the Source Code Of Method in Source Code of Paper (SCOMISCOP) metric measures the proportion of a paper's implementation that is contained within the paper source. We are state-of-the-art for both metrics, outperforming the ResNet and Transformer papers, as well as the PyTorch and Tensorflow libraries. Source code, documentation, videos, crypto scams and an invitation to invest in the commercialisation of NeuRaLaTeX are available at https://www.neuralatex.com

Finlay G. C. Hudson, William A. P. Smith

We present Track Anything Behind Everything (TABE), a novel dataset, pipeline, and evaluation framework for zero-shot amodal completion from visible masks. Unlike existing methods that require pretrained class labels, our approach uses a single query mask from the first frame where the object is visible, enabling flexible, zero-shot inference. Our dataset, TABE-51 provides highly accurate ground truth amodal segmentation masks without the need for human estimation or 3D reconstruction. Our TABE pipeline is specifically designed to handle amodal completion, even in scenarios where objects are completely occluded. We also introduce a specialised evaluation framework that isolates amodal completion performance, free from the influence of traditional visual segmentation metrics.

Ye Yu, Abhimitra Meka, Mohamed Elgharib et al.

Outdoor scene relighting is a challenging problem that requires good understanding of the scene geometry, illumination and albedo. Current techniques are completely supervised, requiring high quality synthetic renderings to train a solution. Such renderings are synthesized using priors learned from limited data. In contrast, we propose a self-supervised approach for relighting. Our approach is trained only on corpora of images collected from the internet without any user-supervision. This virtually endless source of training data allows training a general relighting solution. Our approach first decomposes an image into its albedo, geometry and illumination. A novel relighting is then produced by modifying the illumination parameters. Our solution capture shadow using a dedicated shadow prediction map, and does not rely on accurate geometry estimation. We evaluate our technique subjectively and objectively using a new dataset with ground-truth relighting. Results show the ability of our technique to produce photo-realistic and physically plausible results, that generalizes to unseen scenes.

Ye Yu, William A. P. Smith

In this paper we show how to perform scene-level inverse rendering to recover shape, reflectance and lighting from a single, uncontrolled image using a fully convolutional neural network. The network takes an RGB image as input, regresses albedo, shadow and normal maps from which we infer least squares optimal spherical harmonic lighting coefficients. Our network is trained using large uncontrolled multiview and timelapse image collections without ground truth. By incorporating a differentiable renderer, our network can learn from self-supervision. Since the problem is ill-posed we introduce additional supervision. Our key insight is to perform offline multiview stereo (MVS) on images containing rich illumination variation. From the MVS pose and depth maps, we can cross project between overlapping views such that Siamese training can be used to ensure consistent estimation of photometric invariants. MVS depth also provides direct coarse supervision for normal map estimation. We believe this is the first attempt to use MVS supervision for learning inverse rendering. In addition, we learn a statistical natural illumination prior. We evaluate performance on inverse rendering, normal map estimation and intrinsic image decomposition benchmarks.

William A. P. Smith, Alassane Seck, Hannah Dee et al.

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 (BFM) or FLAME and we make the 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 BFM in albedo reconstruction.

Stylianos Ploumpis, Evangelos Ververas, Eimear O' Sullivan et al.

Three-dimensional Morphable Models (3DMMs) are powerful statistical tools for representing the 3D shapes and textures of an object class. Here we present the most complete 3DMM of the human head to date that includes face, cranium, ears, eyes, teeth and tongue. To achieve this, we propose two methods for combining existing 3DMMs of different overlapping head parts: i. use a regressor to complete missing parts of one model using the other, ii. use the Gaussian Process framework to blend covariance matrices from multiple models. Thus we build a new combined face-and-head shape model that blends the variability and facial detail of an existing face model (the LSFM) with the full head modelling capability of an existing head model (the LYHM). Then we construct and fuse a highly-detailed ear model to extend the variation of the ear shape. Eye and eye region models are incorporated into the head model, along with basic models of the teeth, tongue and inner mouth cavity. The new model achieves state-of-the-art performance. We use our model to reconstruct full head representations from single, unconstrained images allowing us to parameterize craniofacial shape and texture, along with the ear shape, eye gaze and eye color.

Bernhard Egger, William A. P. Smith, Ayush Tewari et al.

In this paper, we provide a detailed survey of 3D Morphable Face Models over the 20 years since they were first proposed. The challenges in building and applying these models, namely capture, modeling, image formation, and image analysis, are still active research topics, and we review the state-of-the-art in each of these areas. We also look ahead, identifying unsolved challenges, proposing directions for future research and highlighting the broad range of current and future applications.

Dizhong Zhu, William A. P. Smith

In this paper, we propose a hybrid depth imaging system in which a polarisation camera is augmented by a second image from a standard digital camera. For this modest increase in equipment complexity over conventional shape-from-polarisation, we obtain a number of benefits that enable us to overcome longstanding problems with the polarisation shape cue. The stereo cue provides a depth map which, although coarse, is metrically accurate. This is used as a guide surface for disambiguation of the polarisation surface normal estimates using a higher order graphical model. In turn, these are used to estimate diffuse albedo. By extending a previous shape-from-polarisation method to the perspective case, we show how to compute dense, detailed maps of absolute depth, while retaining a linear formulation. We show that our hybrid method is able to recover dense 3D geometry that is superior to state-of-the-art shape-from-polarisation or two view stereo alone.

Stylianos Ploumpis, Haoyang Wang, Nick Pears et al.

Three-dimensional Morphable Models (3DMMs) are powerful statistical tools for representing the 3D surfaces of an object class. In this context, we identify an interesting question that has previously not received research attention: is it possible to combine two or more 3DMMs that (a) are built using different templates that perhaps only partly overlap, (b) have different representation capabilities and (c) are built from different datasets that may not be publicly-available? In answering this question, we make two contributions. First, we propose two methods for solving this problem: i. use a regressor to complete missing parts of one model using the other, ii. use the Gaussian Process framework to blend covariance matrices from multiple models. Second, as an example application of our approach, we build a new face-and-head shape model that combines the variability and facial detail of the LSFM with the full head modelling of the LYHM. The resulting combined shape model achieves state-of-the-art performance and outperforms existing head models by a large margin. Finally, as an application experiment, we reconstruct full head representations from single, unconstrained images by utilizing our proposed large-scale model in conjunction with the FaceWarehouse blendshapes for handling expressions.

Sarah Alotaibi, William A. P. Smith

We propose a novel biophysical and dichromatic reflectance model that efficiently characterises spectral skin reflectance. We show how to fit the model to multispectral face images enabling high quality estimation of diffuse and specular shading as well as biophysical parameter maps (melanin and haemoglobin). Our method works from a single image without requiring complex controlled lighting setups yet provides quantitatively accurate reconstructions and qualitatively convincing decomposition and editing.

Ye Yu, William A. P. Smith

We show how to train a fully convolutional neural network to perform inverse rendering from a single, uncontrolled image. The network takes an RGB image as input, regresses albedo and normal maps from which we compute lighting coefficients. Our network is trained using large uncontrolled image collections without ground truth. By incorporating a differentiable renderer, our network can learn from self-supervision. Since the problem is ill-posed we introduce additional supervision: 1. We learn a statistical natural illumination prior, 2. Our key insight is to perform offline multiview stereo (MVS) on images containing rich illumination variation. From the MVS pose and depth maps, we can cross project between overlapping views such that Siamese training can be used to ensure consistent estimation of photometric invariants. MVS depth also provides direct coarse supervision for normal map estimation. We believe this is the first attempt to use MVS supervision for learning inverse rendering.

Anil Bas, William A. P. Smith

We generalise Spatial Transformer Networks (STN) by replacing the parametric transformation of a fixed, regular sampling grid with a deformable, statistical shape model which is itself learnt. We call this a Statistical Transformer Network (StaTN). By training a network containing a StaTN end-to-end for a particular task, the network learns the optimal nonrigid alignment of the input data for the task. Moreover, the statistical shape model is learnt with no direct supervision (such as landmarks) and can be reused for other tasks. Besides training for a specific task, we also show that a StaTN can learn a shape model using generic loss functions. This includes a loss inspired by the minimum description length principle in which an appearance model is also learnt from scratch. In this configuration, our model learns an active appearance model and a means to fit the model from scratch with no supervision at all, even identity labels.

Silvia Tozza, William A. P. Smith, Dizhong Zhu et al.

In this paper we present a differential approach to photo-polarimetric shape estimation. We propose several alternative differential constraints based on polarisation and photometric shading information and show how to express them in a unified partial differential system. Our method uses the image ratios technique to combine shading and polarisation information in order to directly reconstruct surface height, without first computing surface normal vectors. Moreover, we are able to remove the non-linearities so that the problem reduces to solving a linear differential problem. We also introduce a new method for estimating a polarisation image from multichannel data and, finally, we show it is possible to estimate the illumination directions in a two source setup, extending the method into an uncalibrated scenario. From a numerical point of view, we use a least-squares formulation of the discrete version of the problem. To the best of our knowledge, this is the first work to consider a unified differential approach to solve photo-polarimetric shape estimation directly for height. Numerical results on synthetic and real-world data confirm the effectiveness of our proposed method.

Anil Bas, Patrik Huber, William A. P. Smith et al.

In this paper, we show how a 3D Morphable Model (i.e. a statistical model of the 3D shape of a class of objects such as faces) can be used to spatially transform input data as a module (a 3DMM-STN) within a convolutional neural network. This is an extension of the original spatial transformer network in that we are able to interpret and normalise 3D pose changes and self-occlusions. The trained localisation part of the network is independently useful since it learns to fit a 3D morphable model to a single image. We show that the localiser can be trained using only simple geometric loss functions on a relatively small dataset yet is able to perform robust normalisation on highly uncontrolled images including occlusion, self-occlusion and large pose changes.

Anil Bas, William A. P. Smith

A face image contains geometric cues in the form of configurational information and contours that can be used to estimate 3D face shape. While it is clear that 3D reconstruction from 2D points is highly ambiguous if no further constraints are enforced, one might expect that the face-space constraint solves this problem. We show that this is not the case and that geometric information is an ambiguous cue. There are two sources for this ambiguity. The first is that, within the space of 3D face shapes, there are flexibility modes that remain when some parts of the face are fixed. The second occurs only under perspective projection and is a result of perspective transformation as camera distance varies. Two different faces, when viewed at different distances, can give rise to the same 2D geometry. To demonstrate these ambiguities, we develop new algorithms for fitting a 3D morphable model to 2D landmarks or contours under either orthographic or perspective projection and show how to compute flexibility modes for both cases. We show that both fitting problems can be posed as a separable nonlinear least squares problem and solved efficiently. We demonstrate both quantitatively and qualitatively that the ambiguity is present in reconstructions from geometric information alone but also in reconstructions from a state-of-the-art CNN-based method.

Alassane Seck, William A. P. Smith, Arnaud Dessein et al.

We present a practical approach to capturing ear-to-ear face models comprising both 3D meshes and intrinsic textures (i.e. diffuse and specular albedo). Our approach is a hybrid of geometric and photometric methods and requires no geometric calibration. Photometric measurements made in a lightstage are used to estimate view dependent high resolution normal maps. We overcome the problem of having a single photometric viewpoint by capturing in multiple poses. We use uncalibrated multiview stereo to estimate a coarse base mesh to which the photometric views are registered. We propose a novel approach to robustly stitching surface normal and intrinsic texture data into a seamless, complete and highly detailed face model. The resulting relightable models provide photorealistic renderings in any view.

Anil Bas, William A. P. Smith, Timo Bolkart et al.

We propose a fully automatic method for fitting a 3D morphable model to single face images in arbitrary pose and lighting. Our approach relies on geometric features (edges and landmarks) and, inspired by the iterated closest point algorithm, is based on computing hard correspondences between model vertices and edge pixels. We demonstrate that this is superior to previous work that uses soft correspondences to form an edge-derived cost surface that is minimised by nonlinear optimisation.