A wide range of face appearance variations can be modeled by using set-based recognition approaches effectively, but computational complexity of current methods is highly dependent on the set and class sizes. This paper introduces new video-based classification methods designed for reducing the required disk space of data samples and speed up the testing process in large-scale face recognition systems. In the proposed method, image sets collected from videos are approximated with kernelized convex hulls and it was shown that it is sufficient to use only the samples that participate in shaping the image set boundaries in this setting. The kernelized support vector data description (SVDD) is used to extract those important samples that form the image set boundaries. Moreover, we show that these kernelized hypersphere models can also be used to approximate image sets for classification purposes. Then, we propose a binary hierarchical decision tree approach to improve the speed of the classification system even more. At last, we introduce a new video database that includes 285 people with 8 videos of each person, since the most popular video data sets used for set-based recognition methods include either a few people, or small number of videos per person. The experimental results on varying sized databases show that the proposed methods greatly improve the testing times of the classification system (we obtained speed-ups to a factor of 20) without a significant drop in accuracies.

The advent of deep learning has pushed medical image analysis to new levels, rapidly replacing more traditional machine learning and computer vision pipelines. However segmenting and labelling anatomical regions remains challenging owing to appearance variations, imaging artifacts, the paucity and variability of annotated data, and the difficulty of fully exploiting domain constraints such as anatomical knowledge about inter-region relationships. We address the last point, improving the network's region-labeling consistency by introducing NonAdjLoss, an adjacency-graph based auxiliary training loss that penalizes outputs containing regions with anatomically-incorrect adjacency relationships. NonAdjLoss supports both fully-supervised training and a semi-supervised extension in which it is applied to unlabeled supplementary training data. The approach substantially reduces segmentation anomalies on the MICCAI-2012, IBSRv2 brain MRI datasets and the Anatomy3 whole body CT dataset, especially when semi-supervised training is included.

An object detector must detect and localize each instance of the object class of interest in the image. Many recent detectors adopt a sliding window approach, reducing the problem to one of deciding whether the detection window currently contains a valid object instance or background. Machine learning based discriminants such as SVM and boosting are typically used for this, often in the form of classifier cascades to allow more rapid rejection of easy negatives. We argue that “one class” methods - ones that focus mainly on modelling the range of the positive class - are a useful alternative to binary discriminants in such applications, particularly in the early stages of the cascade where one-class approaches may allow simpler classifiers and faster rejection. We implement this in the form of a short cascade of efficient nearest-convex-model one-class classifiers, starting with linear distance-to-affine-hyperplane and interior-of-hypersphere classifiers and finishing with kernelized hypersphere classifiers. We show that our methods have very competitive performance on the Faces in the Wild and ESOGU face detection datasets and state-of-the-art performance on the INRIA Person dataset. As predicted, the one-class formulations provide significant reductions in classifier complexity relative to the corresponding two-class ones.

We introduce the hierarchical Markov aspect model (HMAM), a computationally efficient graphical model for densely labeling large remote sensing images with their underlying terrain classes. HMAM resolves local ambiguities efficiently by combining the benefits of quadtree representations and aspect models-the former incorporate multiscale visual features and hierarchical smoothing to provide improved local label consistency, while the latter sharpen the labelings by focusing them on the classes that are most relevant for the broader local image context. The full HMAM model takes a grid of local hierarchical Markov quadtrees over image patches and augments it by incorporating a probabilistic latent semantic analysis aspect model over a larger local image tile at each level of the quadtree forest. Bag-of-word visual features are extracted for each level and patch, and given these, the parent-child transition probabilities from the quadtree and the label probabilities from the tile-level aspect models, an efficient forwards-backwards inference pass allows local posteriors for the class labels to be obtained for each patch. Variational expectation-maximization is then used to train the complete model from either pixel-level or tile-keyword-level labelings. Experiments on a complete TerraSAR-X synthetic aperture radar terrain map with pixel-level ground truth show that HMAM is both accurate and efficient, providing significantly better results than comparable single-scale aspect models with only a modest increase in training and test complexity. Keyword-level training greatly reduces the cost of providing training data with little loss of accuracy relative to pixel-level training.

International audience

International audience

International audience

We introduce a novel method for face recognition from image sets. In our setting each test and training example is a set of images of an individual's face, not just a single image, so recognition decisions need to be based on comparisons of image sets. Methods for this have two main aspects: the models used to represent the individual image sets; and the similarity metric used to compare the models. Here, we represent images as points in a linear or affine feature space and characterize each image set by a convex geometric region (the affine or convex hull) spanned by its feature points. Set dissimilarity is measured by geometric distances (distances of closest approach) between convex models. To reduce the influence of outliers we use robust methods to discard input points that are far from the fitted model. The kernel trick allows the approach to be extended to implicit feature mappings, thus handling complex and nonlinear manifolds of face images. Experiments on two public face datasets show that our proposed methods outperform a number of existing state-of-the-art ones.

In this paper, we present a fast approach to obtain semantic scene segmentation with high precision. We employ a two-stage classifier to label all image pixels. First, we use the regularized logistic regression to combine different appearance-based features and the improved spatial layout of labeling information. In the second stage, we incorporate the local, regional and global cues into a conditional random field model to provide a final segmentation, and a fast max-margin training method is employed to learn the parameters of the model quickly. The comparison experiments on four multi-class image segmentation databases show that our approach can achieve comparable semantic segmentation results and work faster than that of the state-of-the-art approaches.

Making recognition more reliable under uncontrolled lighting conditions is one of the most important challenges for practical face recognition systems. We tackle this by combining the strengths of robust illumination normalization, local texture-based face representations, distance transform based matching, kernel-based feature extraction and multiple feature fusion. Specifically, we make three main contributions: 1) we present a simple and efficient preprocessing chain that eliminates most of the effects of changing illumination while still preserving the essential appearance details that are needed for recognition; 2) we introduce local ternary patterns (LTP), a generalization of the local binary pattern (LBP) local texture descriptor that is more discriminant and less sensitive to noise in uniform regions, and we show that replacing comparisons based on local spatial histograms with a distance transform based similarity metric further improves the performance of LBP/LTP based face recognition; and 3) we further improve robustness by adding Kernel principal component analysis (PCA) feature extraction and incorporating rich local appearance cues from two complementary sources--Gabor wavelets and LBP--showing that the combination is considerably more accurate than either feature set alone. The resulting method provides state-of-the-art performance on three data sets that are widely used for testing recognition under difficult illumination conditions: Extended Yale-B, CAS-PEAL-R1, and Face Recognition Grand Challenge version 2 experiment 4 (FRGC-204). For example, on the challenging FRGC-204 data set it halves the error rate relative to previously published methods, achieving a face verification rate of 88.1% at 0.1% false accept rate. Further experiments show that our preprocessing method outperforms several existing preprocessors for a range of feature sets, data sets and lighting conditions.

Nearest neighbour classifiers and related kernel methods often perform poorly in high dimensional problems because it is infeasible to include enough training samples to cover the class regions densely. In such cases, test samples often fall into gaps between training samples where the nearest neighbours are too distant to be good indicators of class membership. One solution is to project the data onto a discriminative lower dimensional subspace. We propose a gap-resistant nonparametric method for finding such subspaces: first the gaps are filled by building a convex model of the region spanned by each class - we test the affine and convex hulls and the bounding disk of the class training samples - then a set of highly discriminative directions is found by building and decomposing a scatter matrix of weighted displacement vectors from training examples to nearby rival class regions. The weights are chosen to focus attention on narrow margin cases while still allowing more diversity and hence more discriminability than the 1D linear Support Vector Machine (SVM) projection. Experimental results on several face and object recognition datasets show that the method finds effective projections, allowing simple classifiers such as nearest neighbours to work well in the low dimensional reduced space.

In high-dimensional classification problems it is infeasible to include enough training samples to cover the class regions densely. Irregularities in the resulting sparse sample distributions cause local classifiers such as Nearest Neighbors (NN) and kernel methods to have irregular decision boundaries. One solution is to "fill in the holes" by building a convex model of the region spanned by the training samples of each class and classifying examples based on their distances to these approximate models. Methods of this kind based on affine and convex hulls and bounding hyperspheres have already been studied. Here we propose a method based on the bounding hyperdisk of each class - the intersection of the affine hull and the smallest bounding hypersphere of its training samples. We argue that in many cases hyperdisks are preferable to affine and convex hulls and hyperspheres: they bound the classes more tightly than affine hulls or hyperspheres while avoiding much of the sample overfitting and computational complexity that is inherent in high-dimensional convex hulls. We show that the hyperdisk method can be kernelized to provide nonlinear classifiers based on non-Euclidean distance metrics. Experiments on several classification problems show promising results.

Considerable advances have been made in learning to recognize and localize visual object classes. Simple bag-of-feature approaches label each pixel or patch independently. More advanced models attempt to improve the coherence of the labellings by introducing some form of inter-patch coupling: traditional spatial models such as MRF's provide crisper local labellings by exploiting neighbourhood-level couplings, while aspect models such as PLSA and LDA use global relevance estimates (global mixing proportions for the classes appearing in the image) to shape the local choices. We point out that the two approaches are complementary, combining them to produce aspect-based spatial field models that outperform both approaches. We study two spatial models: one based on averaging over forests of minimal spanning trees linking neighboring image regions, the other on an efficient chain-based Expectation Propagation method for regular 8-neighbor Markov random fields. The models can be trained using either patch-level labels or image-level keywords. As input features they use factored observation models combining texture, color and position cues. Experimental results on the MSR Cambridge data sets show that combining spatial and aspect models significantly improves the region-level classification accuracy. In fact our models trained with image-level labels outperform PLSA trained with pixel-level ones.

Conditional Random Fields (CRFs) are an effective tool for a variety of different data segmentation and labeling tasks including visual scene interpretation, which seeks to partition images into their constituent semantic-level regions and assign appropriate class labels to each region. For accurate labeling it is important to capture the global context of the image as well as local information. We in- troduce a CRF based scene labeling model that incorporates both local features and features aggregated over the whole image or large sections of it. Secondly, traditional CRF learning requires fully labeled datasets which can be costly and troublesome to produce. We introduce a method for learning CRFs from datasets with many unlabeled nodes by marginalizing out the unknown labels so that the log-likelihood of the known ones can be maximized by gradient ascent. Loopy Belief Propagation is used to approximate the marginals needed for the gradi- ent and log-likelihood calculations and the Bethe free-energy approximation to the log-likelihood is monitored to control the step size. Our experimental results show that effective models can be learned from fragmentary labelings and that incorporating top-down aggregate features significantly improves the segmenta- tions. The resulting segmentations are compared to the state-of-the-art on three different image datasets.

Some of the most effective recent methods for content-based image classification work by extracting dense or sparse local image descriptors, quantizing them according to a coding rule such as k-means vector quantization, accumulating histograms of the resulting "visual word" codes over the image, and classifying these with a conventional classifier such as an SVM. Large numbers of descriptors and large codebooks are needed for good results and this becomes slow using k-means. We introduce Extremely Randomized Clustering Forests ensembles of randomly created clustering trees and show that these provide more accurate results, much faster training and testing and good resistance to background clutter in several state-of-the-art image classification tasks.

We describe a learning-based method for recovering 3D human body pose from single images and monocular image sequences. Our approach requires neither an explicit body model nor prior labeling of body parts in the image. Instead, it recovers pose by direct nonlinear regression against shape descriptor vectors extracted automatically from image silhouettes. For robustness against local silhouette segmentation errors, silhouette shape is encoded by histogram-of-shape-contexts descriptors. We evaluate several different regression methods: ridge regression, Relevance Vector Machine (RVM) regression, and Support Vector Machine (SVM) regression over both linear and kernel bases. The RVMs provide much sparser regressors without compromising performance, and kernel bases give a small but worthwhile improvement in performance. The loss of depth and limb labeling information often makes the recovery of 3D pose from single silhouettes ambiguous. To handle this, the method is embedded in a novel regressive tracking framework, using dynamics from the previous state estimate together with a learned regression value to disambiguate the pose. We show that the resulting system tracks long sequences stably. For realism and good generalization over a wide range of viewpoints, we train the regressors on images resynthesized from real human motion capture data. The method is demonstrated for several representations of full body pose, both quantitatively on independent but similar test data and qualitatively on real image sequences. Mean angular errors of 4-6 degrees are obtained for a variety of walking motions.

We propose a generative model that codes the geometry and appearance of generic visual object categories as a loose hierarchy of parts, with probabilistic spatial relations linking parts to subparts, soft assignment of subparts to parts, and scale invariant keypoint based local features at the lowest level of the hierarchy. The method is designed to efficiently handle categories containing hundreds of redundant local features, such as those returned by current key-point detectors. This robustness allows it to outperform constellation style models, despite their stronger spatial models. The model is initialized by robust bottom-up voting over location-scale pyramids, and optimized by expectation-maximization. Training is rapid, and objects do not need to be marked in the training images. Experiments on several popular datasets show the method's ability to capture complex natural object classes.

We study the question of feature sets for robust visual object recognition; adopting linear SVM based human detection as a test case. After reviewing existing edge and gradient based descriptors, we show experimentally that grids of histograms of oriented gradient (HOG) descriptors significantly outperform existing feature sets for human detection. We study the influence of each stage of the computation on performance, concluding that fine-scale gradients, fine orientation binning, relatively coarse spatial binning, and high-quality local contrast normalization in overlapping descriptor blocks are all important for good results. The new approach gives near-perfect separation on the original MIT pedestrian database, so we introduce a more challenging dataset containing over 1800 annotated human images with a large range of pose variations and backgrounds.

This article introduces the absolute quadratic complex formed by all lines that intersect the absolute conic. If /spl omega/ denotes the 3 /spl times/ 3 symmetric matrix representing the image of that conic under the action of a camera with projection matrix P, it is shown that /spl omega/ /spl ap/ P/sup ~//spl Omega//sub /spl I.bar//P/sup ~T/ where V is the 3 /spl times/ 6 line projection matrix associated with P and /spl Omega//sub /spl I.bar// is a 6 /spl times/ 6 symmetric matrix of rank 3 representing the absolute quadratic complex. This simple relation between a camera's intrinsic parameters, its projection matrix expressed in a projective coordinate frame, and the metric upgrade separating this frame from a metric one - as respectively captured by the matrices /spl omega/, P/sup ~/ and /spl Omega//sub /spl I.bar// - provides a new framework for autocalibration, particularly well suited to typical digital cameras with rectangular or square pixels since the skew and aspect ratio are decoupled from the other intrinsic parameters in /spl omega/.

Visual codebook based quantization of robust appearance descriptors extracted from local image patches is an effective means of capturing image statistics for texture analysis and scene classification. Codebooks are usually constructed by using a method such as k-means to cluster the descriptor vectors of patches sampled either densely ('textons') or sparsely ('bags of features' based on key-points or salience measures) from a set of training images. This works well for texture analysis in homogeneous images, but the images that arise in natural object recognition tasks have far less uniform statistics. We show that for dense sampling, k-means over-adapts to this, clustering centres almost exclusively around the densest few regions in descriptor space and thus failing to code other informative regions. This gives suboptimal codes that are no better than using randomly selected centres. We describe a scalable acceptance-radius based clusterer that generates better codebooks and study its performance on several image classification tasks. We also show that dense representations outperform equivalent keypoint based ones on these tasks and that SVM or mutual information based feature selection starting from a dense codebook further improves the performance.

We describe a sparse Bayesian regression method for recovering 3D human body motion directly from silhouettes extracted from monocular video sequences. No detailed body shape model is needed, and realism is ensured by training on real human motion capture data. The tracker estimates 3D body pose by using Relevance Vector Machine regression to combine a learned autoregressive dynamical model with robust shape descriptors extracted automatically from image silhouettes. We studied several different combination methods, the most effective being to learn a nonlinear observation-update correction based on joint regression with respect to the predicted state and the observations. We demonstrate the method on a 54-parameter full body pose model, both quantitatively using motion capture based test sequences, and qualitatively on a test video sequence.

A major difficulty for 3D (three-dimensional) human body tracking from monocular image sequences is the near nonobservability of kinematic degrees of freedom that generate motion in depth. For known link (body segment) lengths, the strict nonobservabilities reduce to twofold 'forwards/backwards flipping' ambiguities for each link. These imply 2/sup # links/ formal inverse kinematics solutions for the full model, and hence linked groups of O(2/sup # links/) local minima in the model-image matching cost function. Choosing the wrong minimum leads to rapid mistracking, so for reliable tracking, rapid methods of investigating alternative minima within a group are needed. Previous approaches to this have used generic search methods that do not exploit the specific problem structure. Here, we complement these by using simple kinematic reasoning to enumerate the tree of possible forwards/backwards flips, thus greatly speeding the search within each linked group of minima. Our methods can be used either deterministically, or within stochastic 'jump-diffusion' style search processes. We give experimental results on some challenging monocular human tracking sequences, showing how the new kinematic-flipping based sampling method improves and complements existing ones.

We present a method for recovering 3D human body motion from monocular video sequences using robust image matching, joint limits and non-self-intersection constraints, and a new sample-and-refine search strategy guided by rescaled cost-function covariances. Monocular 3D body tracking is challenging: for reliable tracking at least 30 joint parameters need to be estimated, subject to highly nonlinear physical constraints; the problem is chronically ill conditioned as about 1/3 of the d.o.f. (the depth-related ones) are almost unobservable in any given monocular image; and matching an imperfect, highly flexible self-occluding model to cluttered image features is intrinsically hard. To reduce correspondence ambiguities we use a carefully designed robust matching-cost metric that combines robust optical flow, edge energy, and motion boundaries. Even so, the ambiguity, nonlinearity and non-observability make the parameter-space cost surface multi-modal, unpredictable and ill conditioned, so minimizing it is difficult. We discuss the limitations of CONDENSATION-like samplers, and introduce a novel hybrid search algorithm that combines inflated-covariance-scaled sampling and continuous optimization subject to physical constraints. Experiments on some challenging monocular sequences show that robust cost modelling, joint and self-intersection constraints, and informed sampling are all essential for reliable monocular 3D body tracking.

We introduce 'Joint Feature Distributions', a general statistical framework for feature based multi-image matching that explicitly models the joint probability distributions of corresponding features across several images. Conditioning on feature positions in some of the images gives well-localized distributions for their correspondents in the others, and hence tight likelihood regions for correspondence search. We apply the framework in the simplest case of Gaussian-like distributions over the direct sum (affine images) and tensor product (projective images) of the image coordinates. This produces probabilistic correspondence models that generalize the geometric multi-image matching constraints, roughly speaking by a form of model-averaging over them. These very simple methods predict accurate correspondence likelihood regions for any scene geometry including planar and near-planar scenes, without ill-conditioning or explicit model selection. Small amounts of distortion and non-rigidity are also tolerated. We develop the theory for any number of affine or projective images, explain its relationship to matching tensors, and give results for an initial implementation.

We study the low-level problem of predicting pixel intensities after subpixel image translations. This is a basic subroutine for image warping and super-resolution, and it has a critical influence on the accuracy of subpixel matching by image correlation. Rather than using traditional frequency-space filtering theory or ad hoc interpolators such as splines, we take an empirical approach, finding optimal subpixel interpolation filters by direct numerical optimization over a large set of training examples. The training set is generated by subsampling larger images at different translations, using subsamplers that mimic the spatial response functions of real pixels. We argue that this gives realistic results, and design filters of various different parametric forms under traditional and robust prediction error metrics. We systematically study the performance of the resulting filters, paying particular attention to the influence of the underlying image sampling regime and the effects of aliasing ("jaggies"). We summarize the results and give practical advice for obtaining subpixel accuracy.

We investigate the motions that lead to ambiguous Euclidean scene reconstructions under several common calibration constraints, giving a complete description of such critical motions for: (i) internally calibrated orthographic and perspective cameras; (ii) in two images, for cameras with unknown focal lengths, either different or equal. One aim of the work was to evaluate the potential of modern algebraic geometry tools for rigorously proving properties of vision algorithms, so we use ideal theoretic calculations as well as classical algebra and geometry. We also present numerical experiments showing the effects of near-critical configurations for the varying and fixed focal length methods.

We describe two direct quasilinear methods for camera pose (absolute orientation) and calibration from a single image of 4 or 5 known 3D points. They generalize the 6 point 'Direct Linear Transform' method by incorporating partial prior camera knowledge, while still allowing some unknown calibration parameters to be recovered. Only linear algebra is required, the solution is unique in non-degenerate cases, and additional points can be included for improved stability. Both methods fail for coplanar points, but we give an experimental eigendecomposition based one that handles both planar and nonplanar cases. Our methods use recent polynomial solving technology, and we give a brief summary of this. One of our aims was to try to understand the numerical behaviour of modern polynomial solvers on some relatively simple test cases, with a view to other vision applications.

We introduce a finite difference expansion for closely spaced cameras in projective vision, and use it to derive differential analogues of the finite-displacement projective matching tensors and constraints. The results are simpler, more general and easier to use than Astrom & Heyden's time-derivative based 'continuous time matching constraints'. We suggest how to use the formalism for 'tensor tracking'-propagation of matching relations against a fixed base image along an image sequence. We relate this to non-linear tensor estimators and show how 'unwrapping the optimization loop' along the sequence allows simple 'linear n point' update estimates to converge rapidly to statistically near-optimal, near-consistent tensor estimates as the sequence proceeds. We also give guidelines as to when difference expansion is likely to be worthwhile as compared to a discrete approach.

The author describes a new method for camera autocalibration and scaled Euclidean structure and motion, from three or more views taken by a moving camera with fixed but unknown intrinsic parameters. The motion constancy of these is used to rectify an initial projective reconstruction. Euclidean scene structure is formulated in terms of the absolute quadric-the singular dual 3D quadric (4/spl times/4 rank 3 matrix) giving the Euclidean dot-product between plane normals. This is equivalent to the traditional absolute conic but simpler to use. It encodes both affine and Euclidean structure, and projects very simply to the dual absolute image conic which encodes camera calibration. Requiring the projection to be constant gives a bilinear constraint between the absolute quadric and image conic, from which both can be recovered nonlinearly from m/spl ges/3 images, or quasi-linearly from m/spl ges/4. Calibration and Euclidean structure follow easily. The nonlinear method is stabler, faster, more accurate and more general than the quasi-linear one. It is based on a general constrained optimization technique-sequential quadratic programming-that may well be useful in other vision problems.

This paper describes initial work on a family of projective reconstruction techniques that compute projection matrices directly and linearly from matching tensors estimated from the image data.The approach is based on 'joint image closure relations' -bilinear constraints between matching tensors and projection matrices, that express the fact that the former derive from the latter.The simplest methods use fundamental matrices and epipoles, alternative ones use trilinear tensors.It is possible to treat all of the image data uniformly, without reliance on 'privileged' images or tokens.The underlying theory is discussed, and the performance of the new methods is quantified and compared with that of several existing ones.

This paper describes a family of factorization-based algorithms that recover 3D projective structure and motion from multiple uncalibrated perspective images of 3D points and lines. They can be viewed as generalizations of the Tomasi-Kanade algorithm from affine to fully perspective cameras, and from points to lines. They make no restrictive assumptions about scene or camera geometry, and unlike most existing reconstruction methods they do not rely on 'privileged' points or images. All of the available image data is used, and each feature in each image is treated uniformly. The key to projective factorization is the recovery of a consistent set of projective depths (scale factors) for the image points: this is done using fundamental matrices and epipoles estimated from the image data. We compare the performance of the new techniques with several existing ones, and also describe an approximate factorization method that gives similar results to SVD-based factorization, but runs much more quickly for large problems.

The paper studies the geometry of multi image perspective projection and the matching constraints that this induces on image measurements. The combined image projections define a 3D joint image subspace of the space of combined homogeneous image coordinates. This is a complete projective replica of the 3D world in image coordinates. Its location encodes the imaging geometry and is captured by the 4 index joint image Grassmannian tensor. Projective reconstruction in the joint image is a canonical process requiring only a simple rescaling of image coordinates. Reconstruction in world coordinates amounts to a choice of basis in the joint image. The matching constraints are multilinear tensorial equations in image coordinates that tell whether tokens in different images could be the projections of a single world token. For 2D images of 3D points there are exactly three basic types: the epipolar constraint, A. Shashua's (1995) trilinear one, and a new quadrilinear 4 image one. For images of lines, R. Hartley's (1994) trilinear constraint is the only type. The coefficients of the matching constraints are tensors built directly from the joint image Grassmannian. Their complex algebraic interdependency is captured by quadratic structural simplicity constraints on the Grassmannian.>

Remote sensors such as CCD cameras can be used for a variety of robot sensing tasks, but given restrictions on camera location and imaging geometry, task constraints, and visual occlusion it can be difficult to find viewing positions from which the task can be completed. The complexity of these constraints suggests that automated, quantitative methods of sensor placement are likely to be useful, particularly when the workspace is cluttered and a mobile robot-mounted sensor is being used to increase the sensible region, circumvent occlusions, and so forth. We describe a camera placement planner designed to produce heuristically good static viewing positions for a robot-mounted CCD camera in an experimental workcell. It can be configured to produce viewpoints for a variety of tasks such as workpiece location, inspection and modelling; feedback control by visual servoing; and task progress monitoring. The planner uses a novel probability-based global search technique to optimize a viewpoint evaluation function that heuristically combines task, camera, robot and environmental constraints. The main advances over previous work are the incorporation of kinematic accessibility and collision constraints in the viewpoint evaluation and the introduction of a search technique powerful enough to handle the resulting strong nonlinearities.