RoboDA: A Dataset for Domain Adaptation in Robot Vision

Domain Adaptation (DA) has been used across multiple applications including classification, segmentation, object detection, etc. However, most of the datasets, used to evaluate DA-based methods, are restricted to a single domain or involve costly setup. Specifically for robot vision, noisy real-world environments have always impeded the generalisation of the corresponding datasets. This has led to a dearth of easy-to-use datasets in this niche that can provide challenging DA landscapes for evaluation. This paper introduces RoboDA, a dataset dedicated to DA for robot vision. RoboDA comprises images from the real world and the simulated domains captured by an autonomous mobile robot, along with the corresponding labels. The dataset is easy to use and facilitates easy replicability of the experimental apparatus. The paper highlights the role of RoboDA as a benchmark to filter out good DA methods by comparing the experimental results obtained on RoboDA using both naive knowledge transfer-based and DA-based methods. The paper also compares the impact of using the simulated domain of RoboDA against a visually similar standard dataset for DA methods, wherein the former performs better, buttressing the role of the RoboDA for DA methods. A qualitative analysis of RoboDA highlights its complex data distribution and feature landscape.