QuMaDe: Quick Foreground Mask and Monocular Depth Data Generation

Segmentation of the desired object along with depth estimation is useful in various applications like robotics and autonomous navigation. Any deep learning workflow to estimate monocular depth and segment the desired foreground object in a scene requires significant training data. The data generation process usually involves expensive hardware like RGB-D sensors, laser scanners, or significant manual involvement. Moreover, for every specific foreground object, the data collection process needs to be repeated. This paper presents a novel way to utilize only a small number of readily available png images with transparency for the foreground object and representative background images from the internet and combine them to generate a large dataset for deep learning, utilizing recent monocular depth estimation techniques. To illustrate the effectiveness of the data generation approach, this paper presents a baseline model for depth and foreground mask estimation for detecting cattle on roads using the generated data from the proposed approach. The baseline model exhibits strong generalization to real scenarios. The generated dataset is available for public use.