Multi-Resolution Shared Representative Filtering for
Real-Time Depth Completion

Yu-Ting Wu1     Tzu-Mao Li2     I-Chao Shen3     Hong-Shiang Lin4     Yung-Yu Chuang1
National Taiwan University1     MIT CSAIL2     The University of Tokyo3     FIH Mobile Limited4

Interactive Comparison on Middlebury 2014 Data

Image viewer is borrowed from this page.

All results are generated on a machine with Intel Core i5-7400K at 3.0 GHz, 16-GB of RAM, and an NVIDIA GeForce GTX 2080 Ti graphics card.

Scenes (click on the thumbnail to select the scene comparison shown at the bottom of the page)
Adirondack
input A
input B
Backpack
input A
input B
Bicycle1
input A
input B
Cable
input A
input B
Classroom1
input A
input B
Couch
input A
input B
Flowers
input A
input B
Jadeplant
input A
input B
Mask
input A
input B
Motorcycle
input A
input B
Piano
input A
input B
Pipes
input A
input B
Playroom
input A
input B
Playtable
input A
input B
Recycle
input A
input B
Shelves
input A
input B
Shopvac
input A
input B
Sticks
input A
input B
Storage
input A
input B
Sword1
input A
input B
Sword2
input A
input B
Umbrella
input A
input B
Vintage
input A
input B

Compared Methods
Multi-Res. Joint Bilateral Upsampling (M-JBU) [Richardt et al. 2012]
Fast Bilateral Solver (FBS) [Barron and Poole 2016]
Multi-Res. Shared Representative Filtering (M-SRF) [Our method]

Implementation and Time Budgets
Multi-Res. Joint Bilateral Upsampling [Richardt et al. 2012] and our method are implemented with Unity shaders, running on GPU. We allocated 20 ms. as time budget to complete the depth map. For Fast Bilateral Solver [Barron and Poole 2016], we use the authors' CPU python implementation. It takes about 1200 ~ 2000 ms.

Mean Absolute Error (MAE) and Peak Signal-to-Noise Ratio (PSNR)
Method M-JBU FBS M-SRF
MAE 0.0260 0.0258 0.0204
PSNR 31.93 33.92 36.13