Research_notes

I started Masters program of data science from March 2021 and decided to leave the history I have tried, failed and accomplished.

1. EdgePass: Lightweight Distracted and Drowsiness Driver Ditection System

   Subject: Realtime driver's status detection model with 11 classes (Drive save, Talking, Drowsiness, Distraction, Drinking, Facing behind, Sending message(L/R), Talking(L/R), Radio adjustment)

   First paper-like outcomes from edge computing course project. Proceed with 3 other students at the class and result more light model for detecting driver status.

   Please refer to uploaded paper

2. Unsupervised human video generation

   Subject: From source video and target image, synthesis the appearance of target people to source video

   With the significant result of basemodel (First Order Motion Model, Aliaksandr Siarohin), tried to solve problems of 1) blurring dynamic patterns in clothes 2) mis-generating of front and backward of people. But it holds because of 1) similar methods with existing architecture and 2) feels need to change the way to approach.

   Please refer to this documents

3. Unsupervised landmark detection

   Subject: From videos, find landmarks of the objects in unsupervised methods

   Please refer to this documents

   Very early stage of research (Oct. 2021) and considering some details of research step.

   1. Find landmarks from videos in unsupervised / self-supervised method
   2. Reshape objects from detected landmarks
   3. Reformat 3D shape from 2D images with detected landmarks