SketchGNN: Semantic Sketch Segmentation with Graph Neural Networks

We introduce SketchGNN , a convolutional graph neural network for semantic segmentation and labeling of freehand vector sketches. We treat an input stroke-based sketch as a graph with nodes representing the sampled points along input strokes and edges encoding the stroke structure information. To predict the per-node labels, our SketchGNN uses graph convolution and a static-dynamic branching network architecture to extract the features at three levels, i.e., point-level, stroke-level, and sketch-level. SketchGNN significantly improves the accuracy of the state-of-the-art methods for semantic sketch segmentation (by 11.2% in the pixel-based metric and 18.2% in the component-based metric over a large-scale challenging SPG dataset) and has magnitudes fewer parameters than both image-based and sequence-based methods.

LIMITATIONS OF DESIGN SPACE-BASED INDICATORS FOR EARLY ROBUSTNESS ASSESSMENT

This paper seeks to address the gap between qualitative Robust Design principles and parameter optimization. The former often fails to consider the challenging amount of details in embodiment and configuration design, while the latter is the widely accepted main thrust in traditional Robust Design. The gap is addressed by exploring the value of five quantitative robustness indicators for Design Space Exploration based on variables, objectives and constraints: The set level indicators, Design Space Size and Pareto Set Dispersion, and the point level indicators, Neighbourhood Performance, Failure Rate and Distance to Failure. As a background for the discussion of the limitations of these indicators an industrial case is presented. The case is an incremental encoder and includes two configurations for comparison, five objectives, eight variables, and a range of constraints. The design spaces are sampled and they show conflicting objectives, dispersed spaces and variables dependencies. Based on this it is suggested that set level indicators are more suitable than point level indicators of early robustness evaluation, but the available indicators are limited in their considerations of design space discontinuity and conflicts.

Land use uncertainty in transportation forecast

Using an integrated land use and travel model system implemented for the Puget Sound region in Washington state, a Bayesian Melding technique is applied to represent variations in land use outcomes, and is propagated into travel choices across a multi-year agent-based simulation. A scenario is considered where zoned capacity is increased around light rail stations. Samples are drawn from the posterior distribution of households to generate travel model inputs. They allow for propagation of land use uncertainty into travel choices, which are themselves assessed for uncertainty by comparing against observed data. Resulting travel measures of zonal vehicle miles traveled (VMT) per capita and light rail station boardings indicate the importance of comparing distributions rather than point forecasts. Results suggest decreased VMT per capita in zones near light rail stations and increased boardings at certain stations with existing development, and less significant impacts around stations with lower initial development capacity. In many cases, individual point level comparisons of scenarios would lead to very different conclusions. Altogether, this finding adds to a line of work demonstrating the policy value of incorporating uncertainty in integrated models and provides a method for assessing these variations in a systematic way.

Neighborhood Context and Juvenile Recidivism: A Spatial Analysis of Organizations and Reoffending Risk

Leveraging point-level spatial data from the Phoenix area, we consider the role of nearby organizations as contextual factors that amplify or reduce reoffending risk among juvenile offenders after court completion. Using survival models, we examine whether residential proximity to seven types of organizations impacts risk of recidivism, net of neighborhood disadvantage and offender characteristics. Aggregate neighborhood disadvantage was not associated with reoffending risk and organizational findings were mixed. Low-level offenders with more total organizations nearby had a higher risk of new property offenses, while the risk of drug and violent reoffending nearly doubled for diversion youth residing near police facilities or detention centers. Individual demographics and prior offense histories remained the strongest, most consistent predictors of juvenile recidivism.

Building an Earth Sytem Model emulator for local monthly temperature

<p>Emulators are computationally cheap statistical devices that derive simplified relationships from otherwise complex climate models. A recently developed Earth System Model (ESM) emulator, MESMER (Beusch et al. 2020), uses a combination of pattern scaling and a variability emulator to emulate ESM initial-condition ensembles. Linear scaling provides the spatially resolved yearly temperature trend projections from global mean temperature trend values. In addition, the variability emulator stochastically models spatio-temporally correlated local variability, yielding a convincing imitation of the internal climate variability displayed within a multi-model initial condition ensemble. The work presented here extends MESMER’s framework to have a monthly downscaling module, so as to provide spatially resolved monthly temperature values from spatially resolved yearly temperature values. For this purpose, a harmonic model is trained on monthly ESM output to capture monthly cycles and their evolution with changing temperature. Once the mean monthly cycle is sufficiently emulated, a process based understanding of the biases within the harmonic model is undertaken. Such entails employing a Gradient Boosting Regressor tree model (GBR) to explain the residuals from the harmonic model using biophysical climate variables such as albedo and thermal fluxes as explanatory variables. These variables can be rated according to their explanatory power when categorising residuals which furthermore elucidates the main physical processes driving biases in the harmonic model within seasons at the grid point level. Finally we add residual variability ontop of the harmonic model outputs to provide convincing imitations of ESM monthly temperature realisations. The residual variability is generated using an AR(1) process coupled to a multivariate trans-gaussian process so as to maintain spatio-temporal correlations and the non-stationarity in monthly variability with increasing yearly temperatures.</p><p>Beusch, L., Gudmundsson, L., & Seneviratne, S. I. (2020). Emulating Earth System Model temperatures: from global mean temperature trajectories to grid-point level realizations on land. Earth System Dynamics, 11(1), 139–159. https://doi.org/10.5194/esd-11-139-2020</p><p> </p><p> </p>

Investigating the most important aspect of elite grass court tennis: Short points

Research has shown that short points (points of 0–4 shots) are crucial in determining the outcome of elite men’s and women’s grass court tennis matches. However, research has not explored the importance of short points in more detail to inform practice design. This study aimed to establish the prevalence and importance of individual rally lengths within short points (i.e. points of 0, 1, 2, 3 and 4 shots) in terms of winning elite grass court tennis matches. Using the recently-validated PWOL ( Percentage of matches in which the Winner Outscored the Loser) method, point-level data from 211 men’s and 209 women’s Wimbledon singles matches between 2015 and 2017 were analysed, with short points stratified into individual rally lengths. Results revealed that 1 shot (aces and missed serve-returns) was the most common rally length, with 0 shots (double faults) the least common. Points won of 1 shot, 2 shots and 4 shots were associated with winning matches and can therefore be considered important, but points won of 0 shots and 3 shots were not associated with match outcome. These results highlight the importance of serving and returning strategies at Wimbledon, and indicate that serves and serve-returns should be afforded focus during grass court training. However, the findings appear to contravene anecdotal assertions that ‘serve plus one’ strategies ( points won of 3 shots) are crucial in elite tennis, as they did not differentiate winning and losing players; so coaches should consider the associated practice designs and amount of time afforded to such strategies.

Weakly Supervised Underwater Fish Segmentation Using Affinity LCFCN

Estimating fish body measurements like length, width, and mass has received considerable research due to its potential in boosting productivity in marine and aquaculture applications. Some methods are based on manual collection of these measurements using tools like a ruler which is time consuming and labour intensive. Others rely on fully-supervised segmentation models to automatically acquire these measurements but require collecting per-pixel labels which are also time consuming. It can take up to 2 minutes per fish to acquire accurate segmentation labels. To address this problem, we propose a segmentation model that can efficiently train on images labeled with point-level supervision, where each fish is annotated with a single click. This labeling scheme takes an average of only 1 second per fish. Our model uses a fully convolutional neural network with one branch that outputs per-pixel scores and another that outputs an affinity matrix. These two outputs are aggregated using a random walk to get the final, refined per-pixel output. The whole model is trained end-to-end using the LCFCN loss and thus we call our method Affinity-LCFCN (A-LCFCN). We conduct experiments on the DeepFish dataset, which contains several fish habitats from north-eastern Australia. The results show that A-LCFCN outperforms a fully-supervised segmentation model when the annotation budget is fixed. They also show that A-LCFCN achieves better segmentation results than LCFCN and a standard baseline.