Brendan McCane

On Benchmarking Optical Flow

Evaluating the performance of optical flow algorithms has been difficult because of the lack of ground truth data sets for complex scenes. We present a new method for generating motion fields from real sequences containing polyhedral objects and present a test suite for benchmarking optical flow algorithms consisting of complex synthetic sequences and real scenes with ground truth. We provide a preliminary quantitative evaluation of seven optical flow algorithms using these synthetic and real sequences. Ultimately, we feel that researchers should benchmark their own algorithms using a standard suite. To that end, we offer our Web site as a repository for standard sequences and results.

Recovering Motion Fields: An Evaluation of Eight Optical Flow Algorithms

Evaluating the performance of optical flow algorithms has been difficult because of the lack of ground-truth data sets for complex scenes. We describe a simple modification to a ray tracer that allows us to generate ground-truth motion fields for scenes of arbitrary complexity. The resulting flow maps are used to assist in the comparison of eight optical flow algorithms using three complex, synthetic scenes. Our study found that a modified version of Lucas and Kanade’s algorithm has superior performance but produces sparse flow maps. Proesmans et al.’s algorithm performs slightly worse, on average, but produces a very dense depth map.

SIFT and SURF Performance Evaluation against Various Image Deformations on Benchmark Dataset

Scene classification in indoor and outdoor environments is a fundamental problem to the vision and robotics community. Scene classification benefits from image features which are invariant to image transformations such as rotation, illumination, scale, viewpoint, noise etc. Selecting suitable features that exhibit such invariances plays a key part in classification performance. This paper summarizes the performance of two robust feature detection algorithms namely Scale Invariant Feature Transform (SIFT) and Speeded up Robust Features (SURF) on several classification datasets. In this paper, we have proposed three shorter SIFT descriptors. Results show that the proposed 64D and 96D SIFT descriptors perform as well as traditional 128D SIFT descriptors for image matching at a significantly reduced computational cost. SURF has also been observed to give good classification results on different datasets.

Lumbar segmental instability: a criterion-related validity study of manual therapy assessment

BackgroundMusculoskeletal physiotherapists routinely assess lumbar segmental motion during the clinical examination of a patient with low back pain. The validity of manual assessment of segmental motion has not, however, been adequately investigated.MethodsIn this prospective, multi-centre, pragmatic, diagnostic validity study, 138 consecutive patients with recurrent or chronic low back pain (R/CLBP) were recruited. Physiotherapists with post-graduate training in manual therapy performed passive accessory intervertebral motion tests (PAIVMs) and passive physiological intervertebral motion tests (PPIVMs). Consenting patients were referred for flexion-extension radiographs. Sagittal angular rotation and sagittal translation of each lumbar spinal motion segment was measured from these radiographs, and compared to a reference range derived from a study of 30 asymptomatic volunteers. Motion beyond two standard deviations from the reference mean was considered diagnostic of rotational lumbar segmental instability (LSI) and translational LSI. Accuracy and validity of the clinical assessments were expressed using sensitivity, specificity, and likelihood ratio statistics with 95% confidence intervals (CI).ResultsOnly translation LSI was found to be significantly associated with R/CLBP (p < 0.05). PAIVMs were specific for the diagnosis of translation LSI (specificity 89%, CI 83–93%), but showed poor sensitivity (29%, CI 14–50%). A positive test results in a likelihood ratio (LR+) of 2.52 (95% CI 1.15–5.53). Flexion PPIVMs were highly specific for the diagnosis of translation LSI (specificity 99.5%; CI 97–100%), but showed very poor sensitivity (5%; CI 1–22%). Likelihood ratio statistics for flexion PPIVMs were not statistically significant. Extension PPIVMs performed better than flexion PPIVMs, with slightly higher sensitivity (16%; CI 6–38%) resulting in a likelihood ratio for a positive test of 7.1 (95% CI 1.7 to 29.2) for translation LSI.ConclusionThis study provides the first evidence reporting the concurrent validity of manual tests for the detection of abnormal sagittal planar motion. PAIVMs and PPIVMs are highly specific, but not sensitive, for the detection of translation LSI. Likelihood ratios resulting from positive test results were only moderate. This research indicates that manual clinical examination procedures have moderate validity for detecting segmental motion abnormality.

Lumbar segmental mobility disorders: comparison of two methods of defining abnormal displacement kinematics in a cohort of patients with non-specific mechanical low back pain

BackgroundLumbar segmental rigidity (LSR) and lumbar segmental instability (LSI) are believed to be associated with low back pain (LBP), and identification of these disorders is believed to be useful for directing intervention choices. Previous studies have focussed on lumbar segmental rotation and translation, but have used widely varying methodologies. Cut-off points for the diagnosis of LSR & LSI are largely arbitrary. Prevalence of these lumbar segmental mobility disorders (LSMDs) in a non-surgical, primary care LBP population has not been established.MethodsA cohort of 138 consecutive patients with recurrent or chronic low back pain (RCLBP) were recruited in this prospective, pragmatic, multi-centre study. Consenting patients completed pain and disability rating instruments, and were referred for flexion-extension radiographs. Sagittal angular rotation and sagittal translation of each lumbar spinal motion segment was measured from the radiographs, and compared to a reference range derived from a study of 30 asymptomatic volunteers. In order to define reference intervals for normal motion, and define LSR and LSI, we approached the kinematic data using two different models. The first model used a conventional Gaussian definition, with motion beyond two standard deviations (2sd) from the reference mean at each segment considered diagnostic of rotational LSMD and translational LSMD. The second model used a novel normalised within-subjects approach, based on mean normalised contribution-to-total-lumbar-motion. An LSMD was then defined as present in any segment that contributed motion beyond 2sd from the reference mean contribution-to-normalised-total-lumbar-motion. We described reference intervals for normal segmental mobility, prevalence of LSMDs under each model, and the association of LSMDs with pain and disability.ResultsWith the exception of the conventional Gaussian definition of rotational LSI, LSMDs were found in statistically significant prevalences in patients with RCLBP. Prevalences at both the segmental and patient level were generally higher using the normalised within-subjects model (2.8 to 16.8% of segments; 23.3 to 35.5% of individuals) compared to the conventional Gaussian model (0 to 15.8%; 4.7 to 19.6%). LSMDs are associated with presence of LBP, however LSMDs do not appear to be strongly associated with higher levels of pain or disability compared to other forms of non-specific LBP.ConclusionLSMDs are a valid means of defining sub-groups within non-specific LBP, in a conservative care population of patients with RCLBP. Prevalence was higher using the normalised within-subjects contribution-to-total-lumbar-motion approach.

Algorithmic Fusion for More Robust Feature Tracking

We present a framework for merging the results of independent feature-based motion trackers using a classification based approach. We demonstrate the efficacy of the framework using corner trackers as an example. The major problem with such systems is generating ground truth data for training. We show how synthetic data can be used effectively to overcome this problem. Our combined system performs better in both dropouts and errors than a correspondence tracker, and had less than half the dropouts at the cost of moderate increase in error compared to a relaxation tracker.

Distance functions for categorical and mixed variables

In this paper, we compare three different measures for computing Mahalanobis-type distances between random variables consisting of several categorical dimensions or mixed categorical and numeric dimensions - regular simplex, tensor product space, and symbolic covariance. The tensor product space and symbolic covariance distances are new contributions. We test the methods on two application domains - classification and principal components analysis. We find that the tensor product space distance is impractical with most problems. Over all, the regular simplex method is the most successful in both domains, but the symbolic covariance method has several advantages including time and space efficiency, applicability to different contexts, and theoretical neatness.

Deep networks are effective encoders of periodicity.

We present a comparative theoretical analysis of representation in artificial neural networks with two extreme architectures, a shallow wide network and a deep narrow network, devised to maximally decouple their representative power due to layer width and network depth. We show that, given a specific activation function, models with comparable VC-dimension are required to guarantee zero error modeling of real functions over a binary input. However, functions that exhibit repeating patterns can be encoded much more efficiently in the deep representation, resulting in significant reduction in complexity. This paper provides some initial theoretical evidence of when and how depth can be extremely effective.

Better than SIFT

Independent evaluation of the performance of feature descriptors is an important part of the process of developing better computer vision systems. In this paper, we compare the performance of several state-of-the art image descriptors including several recent binary descriptors. We test the descriptors on an image recognition application and a feature matching application. Our study includes several recently proposed methods and, despite claims to the contrary, we find that SIFT is still the most accurate performer in both application settings. We also find that general purpose binary descriptors are not ideal for image recognition applications but perform adequately in a feature matching application.

Red blood cell segmentation from SEM images

We present a model based contour tracing approach to the problem of automatically segmenting a Scanning Electron Microscope image of red blood cells. We use a second order polynomial model and a simple Bayesian approach to ensure smooth boundaries, and a postprocess ellipse fitting procedure to cull noise contours. Of all contours detected, 95.7% are correct, with a 0.6% false negative rate, and 4.3% false positive rate on 100 sample images involving more than 11000 red blood cells.