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.

We describe an integrated, real-time multi-camera surveillance system that is able to find and track individuals, acquire and archive facial image sequences, and perform face recognition. The system is based around an inference engine that can extract high-level information from an observed scene, and generate appropriate commands for a set of pan-tilt-zoom (PTZ) cameras. The incorporation of a reliable facial recognition into the high-level feedback is a main novelty of our work, showing how high-level understanding of a scene can be used to deploy PTZ sensing resources effectively. The system comprises a distributed camera system using SQL tables as virtual communication channels, Situation Graph Trees for knowledge representation, inference and high-level camera control, and a variety of visual processing algorithms including an on-line acquisition of facial images, and on-line recognition of faces by comparing image sets using subspace distance. We provide an extensive evaluation of this method using our system for both acquisition of training data, and later recognition. A set of experiments in a surveillance scenario show the effectiveness of our approach and its potential for real applications of cognitive vision.

Understanding and analysing video data from static or mobile surveillance cameras often requires knowledge of the scene and the camera placement. In this article, we provide a way to simplify the user's task of understanding the scene by rendering the camera view as if observed from the user's perspective by estimating his position using a real-time visual SLAM system. Augmenting the view is referred to as hidden view synthesis. Compared to previous work, the current approach improves by simplifying the setup and requiring minimal user input. This is achieved by building a map of the environment using a visual SLAM system and then registering the surveillance camera in this map. By exploiting the map, a different moving camera can render hidden views in real-time at 30Hz. We discuss some of the challenges remaining for full automation. Results are shown in an indoor environment for surveillance applications and outdoors with application to improved safety in transport.

Face recognition in surveillance situations usually requires high resolution face images to be captured from remote active cameras. Since the recognition accuracy is typically a function of the face direction with frontal faces more likely to lead to reliable recognition we propose a system which optimises the capturing of such images by using coarse gaze estimates from a static camera. By considering the potential information gain from observing each target, our system automatically sets the pan, tilt and zoom values (i.e. the field of view) of multiple cameras observing different tracked targets in order to maximise the likelihood of correct identification. The expected gain in information is influenced by the controllable field of view, and by the false positive and negative rates of the identification process, which are in turn a function of the gaze angle. We validate the approach using a combination of simulated situations and real tracking output to demonstrate superior performance over alternative approaches, notably using no gaze information, or using gaze inferred from direction of travel (i.e. assuming each person is always looking directly ahead).We also show results from a live implementation with a static camera and two pan-tilt-zoom devices, involving real-time tracking, processing and control.

In this work we present a consistent probabilistic approach to control multiple, but diverse pan-tilt-zoom cameras concertedly observing a scene. There are disparate goals to this control: the cameras are not only to react to objects moving about, arbitrating conflicting interests of target resolution and trajectory accuracy, they are also to anticipate the appearance of new targets. We base our control function on maximisation of expected mutual information gain, which to our knowledge is novel to the field of computer vision in the context of multiple pan-tilt-zoom camera control. This information theoretic measure yields a utility for each goal and parameter setting, making the use of physical or computational resources comparable. Weighting this utility allows to prioritise certain objectives or targets in the control. The resulting behaviours in typical situations for multi-camera systems, such as camera hand-off, acquisition of close-ups and scene exploration, are emergent but intuitive. We quantitatively show that without the need for hand crafted rules they address the given objectives.

We present an online learning approach for robustly combining unreliable observations from a pedestrian detector to estimate the rough 3D scene geometry from video sequences of a static camera. Our approach is based on an entropy modelling framework, which allows to simultaneously adapt the detector parameters, such that the expected information gain about the scene structure is maximised. As a result, our approach automatically restricts the detector scale range for each image region as the estimation results become more confident, thus improving detector run-time and limiting false positives. 1

Studies support the need for high resolution imagery to identify persons in surveillance videos. However, the use of telephoto lenses sacrifices a wider field of view and thereby increases the uncertainty of other, possibly more interesting events in the scene. Using zoom lenses offers the possibility of enjoying the benefits of both wide field of view and high resolution, but not simultaneously. We approach this problem of balancing these finite imaging resources - or of exploration vs exploitation - using an information-theoretic approach. We argue that the camera parameters - pan, tilt and zoom - should be set to maximise information gain, or equivalently minimising conditional entropy of the scene model, comprised of multiple targets and a yet unobserved one. The information content of the former is supplied directly by the uncertainties computed using a Kalman filter tracker, while the latter is modelled using a rdquobackgroundrdquo Poisson process whose parameters are learned from extended scene observations; together these yield an entropy for the scene. We support our argument with quantitative and qualitative analyses in simulated and real-world environments, demonstrating that this approach yields sensible exploration behaviours in which the camera alternates between obtaining close-up views of the targets while paying attention to the background, especially to areas of known high activity.

The use of a single camera with a zoom lens for tracking involves a continuous arbitration of accuracy vs. reliability. We address this problem with an information-theoretic approach, where we extend zoom selection based on conditional entropy by incorporating the fixation errors into the observation likelihood. We present a thorough analysis of previous approaches, revealing zoom and speed limits, especially how the ratio of process to measurement noise effectively limits the maximally usable zoom for any system tracking with a Kalman filter. This work finally presents means to circumvent aforementioned limitations.