| MS TSA Meeting


Name of the Speaker: Mr. Rohit Choudhary (EE20S002)
Guide: Dr. Kaushik Mitra
Co-Guide: Dr. Mansi Sharma
Online meeting link: https://sites.google.com/view/sharmamansi/
Date/Time: June 26th 2024 (Wednesday), 10:30 AM
Title: Deep Learning Method for Efficient Stereo Depth Estimation and Unpaired Low-Light Image Enhancement. Abstract:

Abstract :

With an increasing demand for advanced computer vision techniques in fields such as autonomous driving, virtual reality, and augmented reality, we explore efficient deep learning methods for stereo depth estimation and low-light image enhancement. Stereo depth estimation relies on stereo matching, where depth is inferred by matching corresponding pixels in stereo image pairs to calculate disparity. Key to stereo matching is the cost volume, representing dissimilarity between pixel pairs, which poses computational challenges. We propose two convolutional neural network-based architectures, SDE-DualENet and 2T-UNet, which circumvent the cost volume construction step and preset maximum disparity parameter dependencies in stereo depth estimation. While the former method eliminates the need for cost-volume, the latter additionally integrates monocular depth information to improve scene geometry prediction, achieving better performance on challenging datasets.

We also address the limitations faced by stereo depth estimation techniques in capturing the intricacies of scenes in high dynamic range imaging. The recent trend in industries and research communities seek to merge 3D and HDR technologies for immersive viewing experiences. Creating cost-effective 3D HDR content requires accurate depth estimation, which is challenging due to SDR cameras' limited sensitivity and complex natural scenes. We address this by proposing a dual-encoder single-decoder convolutional neural network, which excels in learning the disparity while efficiently merging disparity maps obtained from stereo images at varying exposure levels developing an efficient multi-exposure stereo depth estimation framework.

Finally, we address the limitations of supervised low-light image enhancement methods, which require paired datasets of low and normal light images that are difficult to capture in real-world conditions. We present ELEGAN, a lightweight generative adversarial network which employs self-regularized illumination attention mapping and modified residual dense blocks to restore low-light images, reducing noise and enhancing visual quality in real-world scenarios. We evaluate the performance of our proposed methods through qualitative and quantitative comparative analysis with state-of-the-art techniques.