Annotated datasets are an essential ingredient to train, evaluate, compare and productionalize supervised machine learning models.It is therefore imperative that annotations are of high quality.For their creation, good quality management and thereby reliable quality estimates are needed.Then, if quality is insufficient during the annotation process, rectifying measures can be taken to improve it.Quality estimation is often performed by having experts manually label instances as correct or incorrect.But checking all annotated instances tends to be expensive.Therefore, in practice, usually only subsets are inspected; sizes are chosen mostly without justification or regard to statistical power and more often than not, are relatively small.Basing estimates on small sample sizes, however, can lead to imprecise values for the error rate.Using unnecessarily large sample sizes costs money that could be better spent, for instance on more annotations.Therefore, we first describe in detail how to use confidence intervals for finding the minimal sample size needed to estimate the annotation error rate.Then, we propose applying acceptance sampling as an alternative to error rate estimation We show that acceptance sampling can reduce the required sample sizes up to 50% while providing the same statistical guarantees.

Many goalkeeper trainees cannot afford a personal human coach. Hence, they could benefit from a virtual coach that provides personalized feedback about the execution of their training exercises. As a first step towards this goal, we developed an algorithm to detect and classify goalkeeper training exercises using a wearable inertial sensor attached to a goalkeeper glove. We collected data from 14 goalkeeper trainees while performing a series of training exercises (e.g., dives, catches, throws). Our approach first detects the exercises using an event detection algorithm based on a high-pass filter, a peak detector, and Dynamic Time Warping to detect and eliminate irrelevant motion instances. Then, it extracts a set of statistical and heuristic features to describe the different exercises and train a machine learning classifier. Our exercise detection approach retrieves 93.8% of the relevant exercises with 90.6% precision and classifies the detected exercises with an accuracy of 96.5%. The exercises recognized by our algorithm can be used to compute further qualitative metrics about individual exercise executions to provide goalkeepers with relevant feedback about their training.

Despite more than three decades of research in augmented reality and shown positive effect of AR in educational settings, we still don't witness spread of this technology in the schools. Complex technology and limited educational content are among the reasons for this absence. Authoring systems can play a positive role in introduction of AR into the school settings. In this paper we introduce ISAR, a domain-independent authoring system for a camera-projector based interactive tabletop. ISAR allows teachers to create learning content and define interactions for the tabletop themselves, without the need for programming, and hence reduces the entrance barrier of educational practitioners for experimenting with augmented reality and tangible interactions.

Patients of an Anterior Cruciate Ligament (ACL) injury engage in a set of rehabilitation exercises to recover mobility and strength of their injured leg. We developed a smart bandage and user interface that provides live feedback to patients while exercising and computes a series of performance metrics used by orthopedists for assessment of the patient's recovery. The bandage uses smart textile components such as elastic conductive threads and snap buttons. We discuss how the smart textile components contribute to desirable properties of the bandage, such as robustness, washability and user comfort. Furthermore, we present a technique for calibrating the motion sensors on the bandage which is suitable for patients of ACL with limited mobility. Finally, we present the results of a controlled experiment with 10 patients with the goal to assess the accuracy of the bandage's measurements.

In modern showjumping and cross-country riding, the success of the horse-rider-pair is measured by the ability to finish a given course of obstacles without penalties within a given time. A horse performs a successful (penalty-free) jump, if no element of the fence falls during the jump. The success of each jump is determined by the correct take-off point of the horse in front of the fence and the amount of strides a horse does between fences. This paper proposes a solution for tracking gaits and jumps using a smartphone attached to the horse's saddle. We propose an event detection algorithm based on Discrete Wavelet Transform and a peak detection to detect jumps and canter strides between fences. We segment the signal to find gait and jump sections, evaluate statistical and heuristic features and classify the segments using different machine learning algorithms. We show that horse jumps and canter strides are detected with a precision of 94.6% and 89.8% recall. All gaits and jumps are further classified with an accuracy of up to 95.4% and a Kappa coefficient (KC) of up to 93%.