Semi-supervised learning (SSL) has been widely explored in recent years, and it is an effective way of leveraging unlabeled data to reduce the reliance on labeled data. In this work, we adjust neural processes (NPs) to the semi-supervised image classification task, resulting in a new method named NP-Match. NP-Match is suited to this task for two reasons. Firstly, NP-Match implicitly compares data points when making predictions, and as a result, the prediction of each unlabeled data point is affected by the labeled data points that are similar to it, which improves the quality of pseudolabels. Secondly, NP-Match is able to estimate uncertainty that can be used as a tool for selecting unlabeled samples with reliable pseudo-labels. Compared with uncertainty-based SSL methods implemented with Monte Carlo (MC) dropout, NP-Match estimates uncertainty with much less computational overhead, which can save time at both the training and the testing phases. We conducted extensive experiments on four public datasets, and NP-Match outperforms state-of-theart (SOTA) results or achieves competitive results on them, which shows the effectiveness of NPMatch and its potential for SSL.

Convolutional neural networks (CNNs) with log-mel spectrum features have shown promising results for acoustic scene classification tasks. However, the performance of these CNN based classifiers is still lacking as they do not generalise well for unknown environments. To address this issue, we introduce an acoustic spectrum transformation network where traditional log-mel spectrums are transformed into imagined visual features (IVF). The imagined visual features are learned by exploiting the relationship between audio and visual features present in video recordings. An auto-encoder is used to encode images as visual features and a transformation network learns how to generate imagined visual features from log-mel. Our model is trained on a large dataset of Youtube videos. We test our proposed method on the scene classification task of DCASE and ESC-50, where our method outperforms other spectrum features, especially for unseen environments.

We propose a self-supervised method for image relighting of single view images in the wild. The method is based on an auto-encoder which deconstructs an image into two separate encodings, relating to the scene illumination and content, respectively. In order to disentangle this embedding information without supervision, we exploit the assumption that some augmentation operations do not affect the image content and only affect the direction of the light. A novel loss function, called spherical harmonic loss, is introduced that forces the illumination embedding to convert to a spherical harmonic vector. We train our model on large-scale datasets such as Youtube 8M and CelebA. Our experiments show that our method can correctly estimate scene illumination and realistically re-light input images, without any supervision or a prior shape model. Compared to supervised methods, our approach has similar performance and avoids common lighting artifacts.

We present a CNN-based method for outdoor highdynamic-range (HDR) environment map prediction from low-dynamic-range (LDR) portrait images. Our method relies on two different CNN architectures, one for light encoding and another for face-to-light prediction. Outdoor lighting is characterised by an extremely high dynamic range, and thus our encoding splits the environment map data between low and high-intensity components, and encodes them using tailored representations. The combination of both network architectures constitutes an end-to-end method for accurate HDR light prediction from faces at real-time rates, inaccessible for previous methods which focused on low dynamic range lighting or relied on non-linear optimisation schemes. We train our networks using both real and synthetic images, we compare our light encoding with other methods for light representation, and we analyse our results for light prediction on real images. We show that our predicted HDR environment maps can be used as accurate illumination sources for scene renderings, with potential applications in 3D object insertion for augmented reality.

In this paper we present a framework for learning a three layered model of human shape, pose and garment deformation. The proposed deformation model provides intuitive control over the three parameters independently, while producing aesthetically pleasing deformations of both the garment and the human body. The shape and pose deformation layers of the model are trained on a rich dataset of full body 3D scans of human subjects in a variety of poses. The garment deformation layer is trained on animated mesh sequences of dressed actors and relies on a novel technique for human shape and posture estimation under clothing. The key contribution of this paper is that we consider garment deformations as the residual transformations between a naked mesh and the dressed mesh of the same subject.

In this paper we present a framework for generating arbitrary human models and animating them realistically given a few intuitive parameters. Shape and pose space deformation (SPSD) is introduced as a technique for modeling subject specific pose induced deformations from whole-body registered 3D scans. By exploiting examples of different people in multiple poses we are able to realistically animate a novel subject by interpolating and extrapolating in a joint shape and pose parameter space. Our results show that we can produce plausible animations of new people and that greater detail is achieved by incorporating subject specific pose deformations. We demonstrate the application of SPSD to produce subject specific animation sequences driven by RGB-Z performance capture.