Published Paper in ICECCME 2022

      I will be presenting my paper titled "Depth Maps Comparisons from Monocular Images by MiDaS Convolutional Neural Networks and Dense Prediction Transformers" in the International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME) conference held in Maldives in November 2022. A link to my recorded presentation and published paper will be available soon.       This paper was based on the first half of my MSDS Capstone Project - see MSDS tab for more information. The research was a comparative case study on inferring depth maps from monocular images using two Deep Learning frameworks: (1) Convolutional Neural Network (CNN) and (2) Dense Prediction Transformers (DPT). Overall my analysis found the Hybrid-DPT model to exhibit the similar computational efficiency as the CNN (known for their computational efficiency), while maintaining accurate representations of 3D geometries like the Large-DPT model. Accurate representations of 3D geometries was analyzed by taking a statistical measurement for outliers vs inliers, where an outlier was computed by exceeding the average distance of 50 nearby neighbors with a threshold of 2.0 standard deviations.
      Important to note: every image is a 2-dimensional representation of our real 3-dimensional world. Every camera has a pinhole camera model that translates the real world into the image you see. This paper briefly describes reverse engineering that process by taking information from the 3x3 intrinsic camera matrix and forming a 4x4 Q-matrix, which provides pixels on the image (X,Y) a mapping to occupy a virtual 3-dimensional space. In summary, real-world 3D points are captured in a 2D image and reprojected back into a virtual 3D space.