In today’s era, dominated by data, there is an increasing need for meticulously annotated data. Robust and accurately labeled datasets are fundamental cornerstones in refining machine learning (ML) algorithms and increasing the precision of organizational decision-making processes. Their critical role is unequivocal; however, the creation of large and diverse datasets, paired with precise annotations, presents a considerable challenge.
An experimental project
with Carl Zeiss Meditec
For predeveloping prototypes in the field of microsurgery, Carl Zeiss Meditec AG also uses machine learning methods. Here, it is often not the development of code and training model that proves tobe the greatest challenge, but rather the selection and annotation of the training data. In this case, the medical technology experts were faced with an immense number of image files to be evaluated aspart of a new project, which would have taken them over 200 hours of work to annotate on their own.
Industry leaders from Bosch, Zeiss, Hyundai, the MIT-IBM Watson Lab, and Quality Match share their insights and discuss past, present and future challenges in the field of dataset creation for deep learning. The panel addresses the issue of dataset quality as well as future trends.
“GARBAGE IN - GARBAGE OUT” is a saying in Data Science. The better the quality of data sets, the better the accuracy of machine learning models – ultimately leading to better products. As enterprises across all verticals, including automotive, rush to apply artificial intelligence (AI) to their business practices, they stumble upon one major obstacle: efficient data labeling at scale.
We annotate 3D bounding boxes from scratch, or based on existing initial guesses, e.g.generated by annotators or algorithms. In our visualizations, we dynamically link RGB and 3D point clouds so that annotators can work very intuitively.
We deliver 3D bounding box distributions based on multiple annotations which are aggregated with uncertainty estimation. As a result, we can rank all boxes according to their ambiguity, so that you can decide which boxes are good enough for training or validation purposes. Optionally, we also identify annotators struggling with the concept of 3D point clouds.
Quality Match News and Events.