Timothy J. Roberts

Human Pose Estimation Using Learnt Probabilistic Region Similarities and Partial Configurations

A model of human appearance is presented for efficient pose estimation from real-world images. In common with related approaches, a high-level model defines a space of configurations which can be associated with image measurements and thus scored. A search is performed to identify good configuration(s). Such an approach is challenging because the configuration space is high dimensional, the search is global, and the appearance of humans in images is complex due to background clutter, shape uncertainty and texture.

Estimating the motion of plant root cells from in vivo confocal laser scanning microscopy images

Images of cellular structures in growing plant roots acquired using confocal laser scanning microscopy have some unusual properties that make motion estimation challenging. These include multiple motions, non-Gaussian noise and large regions with little spatial structure. In this paper, a method for motion estimation is described that uses a robust multi-frame likelihood model and a technique for estimating uncertainty. An efficient region-based matching approach was used followed by a forward projection method. Over small timescales the dynamics are simple (approximately locally constant) and the change in appearance small. Therefore, a constant local velocity model is used and the MAP estimate of the joint probability over a set of frames is recovered. Occurrences of multiple modes in the posterior are detected, and in the case of a single dominant mode, motion is inferred using Laplace’e method. The method was applied to several Arabidopsis thaliana root growth sequences with varying levels of success. In addition, comparative results are given for three alternative motion estimation approaches, the Kanade–Lucas–Tomasi tracker, Black and Anandan’s robust smoothing method, and Markov random field based methods.

Performance of Low-Level Motion Estimation Methods for Confocal Microscopy of Plant Cells in vivo

The performance of various low-level motion estimation methods applied to fluorescence labelled growing cellular structures imaged using confocal laser scanning microscopy is investigated. This is a challenging and unusual domain for motion estimation methods. A selection of methods are discussed that can be contrasted in terms of how much spatial or temporal contextual information is used. The Lucas Kanade feature tracker, a spatially and temporally localised method, was, as one would expect, accurate around resolvable structure. It was not able to track the smaller, repetitive cell structure in the root tip and was somewhat prone to identifying spurious features. This approach is improved by developing a full multi-frame, robust, Bayesian method, and it is demonstrated that by using extra frames with motion constraints reduces such errors. Next, spatially global methods are discussed, including robust variational smoothing and Markov Random Field (MRF) modelling. A key conclusion that is drawn from investigation of these methods is that generic low-level (robust) smoothing functions do not provide good results in this application and that this is probably due to the large regions with little stable structure. Furthermore, contrary to recently reported successes, graph cuts and loopy belief propagation for MAP estimation of the MRF labels provided often poor and inconsistent estimates. The results suggest the need for greater emphasis on temporal smoothing for generic low-level motion estimation tools and more task specific, spatial constraints, perhaps in the form of high level models in order to accurately recover motion from such data. Finally, the form of the estimated growth is briefly discussed and related to contemporary biological models. We hope that this paper will assist non-specialists in applying state-of-the-art methods to this form of data.

Human Pose Estimation Using Partial Configurations and Probabilistic Regions

A method for recovering a part-based description of human pose from single images of people is described. It is able to perform estimation efficiently in the presence significant background clutter, large foreground variation, self-occlusion and occlusion by other objects. This is achieved through two key developments. Firstly, a new formulation is proposed that allows partial configurations, hypotheses with differing numbers of parts, to be made and compared. This permits efficient global sampling in the presence of self and other object occlusions without prior knowledge of body part visibility. Secondly, a highly discriminatory likelihood model is proposed comprising two complementary components. A boundary component improves upon previous appearance distribution divergence methods by incorporating high-level shape and appearance information and hence better discriminates textured, overlapping body parts. An inter-part component uses appearance similarity of body parts to reduce the number of false-positive, multi-part hypotheses, hence increasing estimation efficiency. Results are presented for challenging images with unknown subject and large variations in subject appearance, scale and pose.

Adaptive Learning of Statistical Appearance Models for 3D Human Tracking

A likelihood formulation for human tracking is presented based upon matching feature statistics on the surface of an articulated 3D body model. A benefi to fsuch a formulation over current techniques is that it provides a dense, object-based cue. Multi-dimensional histograms are used to represent feature distributions and different histogram similarity measures are evaluated. An on-line region grouping algorithm, driven by prior knowledge of clothing structure, is derived that enables better histogram estimation and greatly increases computational efficiency. Finally, we demonstrate that the smooth, broad likelihood response allows efficient inference using coarse sampling and local optimisation. Results from tracking real world sequences are presented.

Part-Based Multi-Frame Registration for Estimation of the Growth Of Cellular Networks in Plant Roots

Motion estimation from confocal scanning laser microscope images of growing plant cell structures presents interesting challenges; motion exhibits multiple local discontinuities and noise is non-isotropic and non-Gaussian. A method is presented for estimating motion of cell networks based on a physically motivated, part-based model of cell boundary structure. Each part models the shape and appearance of a localised image region and can undergo constrained non-rigid deformation. This enables motion discontinuities between parts to be modelled. Parts are coupled in order to improve localisation and increase computational efficiency. Results from applying MCMC show accurate localisation of the structure across multiple frames. The form of the model assists biologists in interpreting growth

Human tracking using 3D surface colour distributions

A likelihood formulation for detailed human tracking in real world scenes is presented. In this formulation, the appearance, modelled using feature distributions defined over regions on the surface of an articulated 3D model, is estimated and propagated as part of the state. The benefit of such a formulation over currently used techniques is that it provides a dense, highly discriminatory object-based cue that applies in real world scenes. Multi-dimensional histograms are used to represent the feature distributions and an on-line clustering algorithm, driven by prior knowledge of clothing structure, is derived that enhances appearance estimation and computational efficiency. An investigation of the likelihood model shows its profile to be smooth and broad while region grouping is shown to improve localisation and discrimination. These properties of the likelihood model ease pose estimation by allowing coarse, hierarchical sampling and local optimisation.