Bringing the research advances in Machine Learning (ML) to production is necessary for businesses to gain value from ML. A key challenge of production ML is the monitoring and management of real-time prediction quality. This is complicated by the variability of live production data, the absence of real-time labels and the non-determinism posed by ML techniques themselves. We define ML Health as the real time assessment of ML prediction quality and present an approach to monitoring and improving ML Health. Specifically, a complete solution to monitor and manage ML Health within a realistic full production ML lifecycle. We describe a number of ML Health techniques and assess their efficacy via publicly available datasets. Our solution handles production realities such as scale, heterogeneity and distributed runtimes. We present what we believe is the first solution to production ML Health explored at both an empirical and complete system implementation level.
Around 3,000,000 million vehicle miles are annually traveled utilizing the US transportation systems alone. In addition to the road traffic safety, maintaining the road infrastructure in a sound condition promotes a more productive and competitive economy. Due to the significant amounts of financial and human resources required to detect surface cracks by visual inspection, detection of these surface defects are often delayed resulting in deferred maintenance operations. This paper introduces an automatic system for acquisition, detection, classification, and evaluation of pavement surface cracks by unsupervised analysis of images collected from a camera mounted on the rear of a moving vehicle. A Hessian-based multi-scale filter has been utilized to detect ridges in these images at various scales. Post-processing on the extracted features has been implemented to produce statistics of length, width, and area covered by cracks, which are crucial for roadway agencies to assess pavement quality. This process has been realized on three sets of roads with different pavement conditions in the city of Brockton, MA. A ground truth dataset labeled manually is made available to evaluate this algorithm and results rendered more than 90% segmentation accuracy demonstrating the feasibility of employing this approach at a larger scale.
Video health monitoring of large road networks requires the repeated collection of surface images to detect the defects
and their changes over time. Vehicle mounted video equipment can easily collect the data, but the amount of data that
can be collected in a single day prohibits interactive or semi-automated processing schemes as they would also not be
cost-effective. A new approach that is fully automated to detect road surface defects from large amounts of highresolution
grayscale images is presented. The images are collected with a vehicle-mounted rear-facing 5MP video
camera complemented by GPS based positioning information. Our algorithm starts by correcting the images for radial
and angular distortion to get a bird's-eye view image. This results in images with known dimensions (consistent in width
per pixel) which allow data to be accurately placed on geo-referenced maps. Each of the pixels in the image is labeled as
crack or non-crack using a Markov Random Field (MRF) approach. The data used for testing and training are disjoint
sets of images collected from the streets of Boston, MA, USA. We compare our road surface defect detection results
with other techniques/algorithms described in the literature for accuracy and robustness.
Conference Committee Involvement (2)
Applications of Machine Learning 2020
23 August 2020 | San Diego, California, United States
Applications of Machine Learning
13 August 2019 | San Diego, California, United States