Recent advances in deep learning relies heavily on availability of huge amount of clean annotated information involving substantial manual effort. This becomes a major challenge for tasks related to an entire scene understanding in contrast to a simple object classification task. Researchers are focusing on efficient techniques to use already available synthetic data from various 3D game engines for training artificial vision systems. But, the key question is: how can we adapt such systems for natural scenes with minimal supervision? In a recent study by Prof. Venkatesh Babu and his team from Video Analytics Lab, Department of Computational and Data Sciences, have proposed a novel adversarial learning approach to efficiently adapt synthetically trained models for real scene images. The proposed adaptation method can be used to disentangle the domain discrepancy for improved performance on real scenes from the deployed environment in an unsupervised manner. Currently, the focus is on the task of depth estimation from monocular RGB image scene. But the approach can be extended to other scene understanding problems as well. The method uses an efficient content consistency regularization along with an adversarial learning objective function to train the base Convolutional Neural Network (CNN) architecture. Moreover, the proposed regularization helps to efficiently maintain the spatial dependency of deep features with respect to the given input during the adaptation process. This work will be presented at this year’s CVPR conference.
Reference: Jogendra Nath Kundu, Phani Krishna, Anuj P. and R. Venkatesh Babu, “AdaDepth: Unsupervised Content Congruent Adaptation for Depth Estimation”, in CVPR 2018