21 Apr 2021
Approach
For this project, our group utilized transfer learning using resnet-18 and Unet as pretrained models. We trained each model using the Dense Indoor and Outdoor Depth (DIODE) dataset. Although, to increase the training time, we resized the images in the dataset from 768x1024 to 384x512. After training, our group created a small python script to simulate a lense focusing on a particular depth.
Indoor Results
From left to right, RGB image, ground-truth depth map, ResNet18 predicted depth map, and UNet predicted depth examples from the DIODE indoor data set.
Outdoor Results
From left to right, RGB image, ground-truth depth map, ResNet18 predicted depth map, and UNet predicted depth examples from the DIODE outdoor data set.