Yongxin Yang

Sketch-a-Net that Beats Humans

We propose a multi-scale multi-channel deep neural network framework that, for the first time, yields sketch recognition performance surpassing that of humans. Our superior performance is a result of explicitly embedding the unique characteristics of sketches in our model: (i) a network architecture designed for sketch rather than natural photo statistics, (ii) a multi-channel generalisation that encodes sequential ordering in the sketching process, and (iii) a multi-scale network ensemble with joint Bayesian fusion that accounts for the different levels of abstraction exhibited in free-hand sketches. We show that state-of-the-art deep networks specifically engineered for photos of natural objects fail to perform well on sketch recognition, regardless whether they are trained using photo or sketch. Our network on the other hand not only delivers the best performance on the largest human sketch dataset to date, but also is small in size making efficient training possible using just CPUs.

Sketch-a-Net: A Deep Neural Network that Beats Humans

We propose a deep learning approach to free-hand sketch recognition that achieves state-of-the-art performance, significantly surpassing that of humans. Our superior performance is a result of modelling and exploiting the unique characteristics of free-hand sketches, i.e., consisting of an ordered set of strokes but lacking visual cues such as colour and texture, being highly iconic and abstract, and exhibiting extremely large appearance variations due to different levels of abstraction and deformation. Specifically, our deep neural network, termed Sketch-a-Net has the following novel components: (i) we propose a network architecture designed for sketch rather than natural photo statistics. (ii) Two novel data augmentation strategies are developed which exploit the unique sketch-domain properties to modify and synthesise sketch training data at multiple abstraction levels. Based on this idea we are able to both significantly increase the volume and diversity of sketches for training, and address the challenge of varying levels of sketching detail commonplace in free-hand sketches. (iii) We explore different network ensemble fusion strategies, including a re-purposed joint Bayesian scheme, to further improve recognition performance. We show that state-of-the-art deep networks specifically engineered for photos of natural objects fail to perform well on sketch recognition, regardless whether they are trained using photos or sketches. Furthermore, through visualising the learned filters, we offer useful insights in to where the superior performance of our network comes from.

Weakly Supervised Learning of Objects, Attributes and Their Associations

When humans describe images they tend to use combinations of nouns and adjectives, corresponding to objects and their associated attributes respectively. To generate such a description automatically, one needs to model objects, attributes and their associations. Conventional methods require strong annotation of object and attribute locations, making them less scalable. In this paper, we model object-attribute associations from weakly labelled images, such as those widely available on media sharing sites (e.g. Flickr), where only image-level labels (either object or attributes) are given, without their locations and associations. This is achieved by introducing a novel weakly supervised non-parametric Bayesian model. Once learned, given a new image, our model can describe the image, including objects, attributes and their associations, as well as their locations and segmentation. Extensive experiments on benchmark datasets demonstrate that our weakly supervised model performs at par with strongly supervised models on tasks such as image description and retrieval based on object-attribute associations.

Transductive Multi-label Zero-shot Learning

Zero-shot learning has received increasing interest as a means to alleviate the often prohibitive expense of annotating training data for large scale recognition problems. These methods have achieved great success via learning intermediate semantic representations in the form of attributes and more recently, semantic word vectors. However, they have thus far been constrained to the single-label case, in contrast to the growing popularity and importance of more realistic multi-label data. In this paper, for the first time, we investigate and formalise a general framework for multi-label zero-shot learning, addressing the unique challenge therein: how to exploit multi-label correlation at test time with no training data for those classes? In particular, we propose (1) a multi-output deep regression model to project an image into a semantic word space, which explicitly exploits the correlations in the intermediate semantic layer of word vectors; (2) a novel zero-shot learning algorithm for multi-label data that exploits the unique compositionality property of semantic word vector representations; and (3) a transductive learning strategy to enable the regression model learned from seen classes to generalise well to unseen classes. Our zero-shot learning experiments on a number of standard multi-label datasets demonstrate that our method outperforms a variety of baselines.

Frankenstein: Learning Deep Face Representations Using Small Data

Deep convolutional neural networks have recently proven extremely effective for difficult face recognition problems in uncontrolled settings. To train such networks, very large training sets are needed with millions of labeled images. For some applications, such as near-infrared (NIR) face recognition, such large training data sets are not publicly available and difficult to collect. In this paper, we propose a method to generate very large training data sets of synthetic images by compositing real face images in a given data set. We show that this method enables to learn models from as few as 10 000 training images, which perform on par with models trained from 500 000 images. Using our approach, we also obtain state-of-the-art results on the CASIA NIR-VIS2.0 heterogeneous face recognition data set.

Deeper, Broader and Artier Domain Generalization

The problem of domain generalization is to learn from multiple training domains, and extract a domain-agnostic model that can then be applied to an unseen domain. Domain generalization (DG) has a clear motivation in contexts where there are target domains with distinct characteristics, yet sparse data for training. For example recognition in sketch images, which are distinctly more abstract and rarer than photos. Nevertheless, DG methods have primarily been evaluated on photo-only benchmarks focusing on alleviating the dataset bias where both problems of domain distinctiveness and data sparsity can be minimal. We argue that these benchmarks are overly straightforward, and show that simple deep learning baselines perform surprisingly well on them. In this paper, we make two main contributions: Firstly, we build upon the favorable domain shift-robust properties of deep learning methods, and develop a low-rank parameterized CNN model for end-to-end DG learning. Secondly, we develop a DG benchmark dataset covering photo, sketch, cartoon and painting domains. This is both more practically relevant, and harder (bigger domain shift) than existing benchmarks. The results show that our method outperforms existing DG alternatives, and our dataset provides a more significant DG challenge to drive future research.

Zero-Shot Domain Adaptation via Kernel Regression on the Grassmannian

Most visual recognition methods implicitly assume the data distribution remains unchanged from training to testing. However, in practice domain shift often exists, where real-world factors such as lighting and sensor type change between train and test, and classifiers do not generalise from source to target domains. It is impractical to train separate models for all possible situations because collecting and labelling the data is expensive. Domain adaptation algorithms aim to ameliorate domain shift, allowing a model trained on a source to perform well on a different target domain. However, even for the setting of unsupervised domain adaptation, where the target domain is unlabelled, collecting data for every possible target domain is still costly. In this paper, we propose a new domain adaptation method that has no need to access either data or labels of the target domain when it can be described by a parametrised vector and there exits several related source domains within the same parametric space. It greatly reduces the burden of data collection and annotation, and our experiments show some promising results.

Weakly-Supervised Image Annotation and Segmentation with Objects and Attributes

We propose to model complex visual scenes using a non-parametric Bayesian model learned from weakly labelled images abundant on media sharing sites such as Flickr. Given weak image-level annotations of objects and attributes without locations or associations between them, our model aims to learn the appearance of object and attribute classes as well as their association on each object instance. Once learned, given an image, our model can be deployed to tackle a number of vision problems in a joint and coherent manner, including recognising objects in the scene (automatic object annotation), describing objects using their attributes (attribute prediction and association), and localising and delineating the objects (object detection and semantic segmentation). This is achieved by developing a novel Weakly Supervised Markov Random Field Stacked Indian Buffet Process (WS-MRF-SIBP) that models objects and attributes as latent factors and explicitly captures their correlations within and across superpixels. Extensive experiments on benchmark datasets demonstrate that our weakly supervised model significantly outperforms weakly supervised alternatives and is often comparable with existing strongly supervised models on a variety of tasks including semantic segmentation, automatic image annotation and retrieval based on object-attribute associations.

Spectroscopic super-resolution fluorescence cell imaging using ultra-small Ge quantum dots

We demonstrate a spectroscopic imaging based super-resolution approach by separating the overlapping diffraction spots into several detectors during a single scanning period and taking advantage of the size-dependent emission wavelength in nanoparticles. This approach has been tested using off-the-shelf quantum dots (Invitrogen Qdot) and in-house novel ultra-small (~3 nm) Ge QDs. Furthermore, we developed a method-specific Gaussian fitting and maximum likelihood estimation based on a Matlab algorithm for fast QD localisation. This methodology results in a three-fold improvement in the number of localised QDs compared to non-spectroscopic images. With the addition of advanced ultra-small Ge probes, the number can be improved even further, giving at least 1.5 times improvement when compared to Qdots. Using a standard scanning confocal microscope we achieved a data acquisition rate of 200 ms per image frame. This is an improvement on single molecule localisation super-resolution microscopy where repeated image capture limits the imaging speed, and the size of fluorescence probes limits the possible theoretical localisation resolution. We show that our spectral deconvolution approach has a potential to deliver data acquisition rates on the ms scale thus providing super-resolution in live systems.

Multivariate Regression on the Grassmannian for Predicting Novel Domains

We study the problem of predicting how to recognise visual objects in novel domains with neither labelled nor unlabelled training data. Domain adaptation is now an established research area due to its value in ameliorating the issue of domain shift between train and test data. However, it is conventionally assumed that domains are discrete entities, and that at least unlabelled data is provided in testing domains. In this paper, we consider the case where domains are parametrised by a vector of continuous values (e.g., time, lighting or view angle). We aim to use such domain metadata to predict novel domains for recognition. This allows a recognition model to be pre-calibrated for a new domain in advance (e.g., future time or view angle) without waiting for data collection and re-training. We achieve this by posing the problem as one of multivariate regression on the Grassmannian, where we regress a domains subspace (point on the Grassmannian) against an independent vector of domain parameters. We derive two novel methodologies to achieve this challenging task: a direct kernel regression from RM ! G, and an indirect method with better extrapolation properties. We evaluate our methods on two crossdomain visual recognition benchmarks, where they perform close to the upper bound of full data domain adaptation. This demonstrates that data is not necessary for domain adaptation if a domain can be parametrically described.