Investigating Biases in Visual Navigation Datasets

Abstract

Deep learning has made exciting advances is multiple domains. One major reason of success is the availability of big data. However, gathering lot of data with human e ort introduces lot of biases in the dataset. The recogni- tion community has started investigating these biases. This has helped them to mitigate the a ect of such biases on the recognition system. In this work we plan to investigate such biases in visual navigation datasets. Visual navigation datasets involve specialized sensors (sensors, lasers, stereo cameras). It is di cult to crowd source data for these sensors as com- pared to simple monocular camera. Therefore, datasets are gathered under the supervision of only few individuals. This lack of diversity in the collection stage might result is stronger biases in the visual navigation datasets. We plan to investigate these biases. Furthermore, realistic simulators are gaining traction recently. We plan to investigate the recent realistic simulators and determine their simulation to reality gap. This investigation is required to build visual navigation models that will assist in developing predictable navigation methods for autonomous agents beyond lab settings.