Denoising approaches for data preparation in machine learning

Abstract

Machine-learning models have been recently developed in various computer vision applications, such as image detention, segmentation and so on. Models are often trained based on data that present various levels of accuracy, with typically a better quality of data leading to a higher accuracy of prediction.

This study addresses the importance of training dataset preparation for machine learning applications. We propose to fuse different data modalities for generating our dataset for training a model. However, data streams are not immune from noise and are often desynchronised. Directly employing a noisy training dataset leads to an inaccurate model and predictions. As an alternative to seeking solutions by creating more complex models to tackle this issue, we look instead at improving the quality of the training datasets to train better models.

Our primary contribution is to diminish the noise in the training dataset. We show experimentally that our approach improves performance in applications to building height estimation from single aerial imagery and geotagging of objects from street view images.

For predicting building height from aerial images, our model manages to be within 1.5 meters of the ground truth thanks to a training dataset fusing information from point cloud data and aerial imagery. For geotagging objects, the object of interest can be Geo-localised with higher accuracy by correcting the metadata associated with street view images.

These results suggest that the quality of training examples significantly impacts the result of a model prediction, which leads to a better performance.