Comparing the robustness of different depth map algorithms … · 2019. 3. 8. · •Method 1:...

3
Comparing the robustness of different depth map algorithms using light fields Linda Banh Fang-Yu Lin Motivation: To improve 3D reconstruction techniques for occlusion in augmented reality (AR). Improved depth map techniques could lead to higher accuracy in creating 3D scenes and improving 3D maps of the world for AR. Comparing the robustness of different techniques could advance research in this area.

Transcript of Comparing the robustness of different depth map algorithms … · 2019. 3. 8. · •Method 1:...

Page 1: Comparing the robustness of different depth map algorithms … · 2019. 3. 8. · •Method 1: Using focal stack shift/add •MSE = 0.044, Computation time (tic toc) = 230 s à3.83

Comparing the robustness of different depth map algorithms

using light fieldsLinda Banh Fang-Yu Lin

Motivation: To improve 3D reconstruction techniques for occlusion inaugmented reality (AR). Improved depth map techniques could lead tohigher accuracy in creating 3D scenes and improving 3D maps of theworld for AR. Comparing the robustness of different techniques couldadvance research in this area.

Page 2: Comparing the robustness of different depth map algorithms … · 2019. 3. 8. · •Method 1: Using focal stack shift/add •MSE = 0.044, Computation time (tic toc) = 230 s à3.83

Goal and Methodology

Goal: To compare different depth map algorithms using light fields and to see how it performs when reconstructing 3D scenes.

Plan of Action1. Use the light field images from 4D Light Field Benchmark for

ground truth depth maps2. Use 3 different algorithms – focal stack gradient, epipolar planes,

line fitting – to extract depth map3. Evaluate mean squared error of each depth map with ground truth

views and computation time

Page 3: Comparing the robustness of different depth map algorithms … · 2019. 3. 8. · •Method 1: Using focal stack shift/add •MSE = 0.044, Computation time (tic toc) = 230 s à3.83

Dataset and Initial Results

• 4D Light Field Benchmark (Honauer, Katrin 2016)

• Light field images with ground truth depth map

Ground Truth Result Method 1:

Depth Map

• Use Lytro image with a shift & add algorithm to artificially create different focused images

• Method 1: Using focal stack shift/add

• MSE = 0.044, Computation time (tic toc) = 230 s à 3.83 min

Result Method 1:

All-in-focus

reconstruction