Exploration of open source photogrammetry

About me

CTO & Co-Founder of Obrary

Focused on Product Visualization and Manufacturing automation

Out product is a sales experience SDK for brands, retailers and manufacturers of personalized & customized products

Our interest in VR

VR provides new opportunities to create interactive/engaging sales experiences for products

Extreme interest in Reality Capture & WebVR

Research Objective(s)

Q: can we develop a cost effective workflow for capturing products.

Textures quality is critical
Wide range of object:
- Shapes
- Sizes
- Reflectance

Questions:

Are there open source solutions that we can use commercially?
Are open source solutions mature enough to use?
Can a fixed time investment to capture session ensure success?
What Type of Capture rigs might we have to design/build?
How much compute (time) is required to process
What operator skills are required

Next: What Technologies?

OPTIONS for Reality Capture - RGB-D (IR)

($$) Kinect
($$) Structure (Occipital)
($$$$$$) ArcTec EVA
($$) Tango (tablet/phone) *

*If you have a tango device, i would like to chat with you.

OPTIONS FOR REALITY CAPTURE - SL/Laser

($$$$-$$$$$) David Laser
($$) Eora-3d
($) DIY Laser

OPTIONS FOR REALITY CAPTURE - Photogrammetry

Commercial*

AutoDesk ReCap/Memento/ReMake
Agisoft PhotoScan
Capturing Reality(.com)
Pix4D

Open Source*

VisualSFM /CMMS + MeshLab
VisualSFM/OpenMVS

*For an more comprehensive list:

https://en.wikipedia.org/wiki/Comparison_of_photogrammetry_software

The focus of this talk

Why Photogrammetry?

Easy to "overcapture"
Reality textures
Quality textures
Software solution
Many vendors (Free/Paid)
Not everything has CAD file(s)
Realistic Aesthetic

The chief objective of a photogrammetric process is to determine the distance between points in 3d space.

Simple Right?

Basics

Point - simple a point in 3D space (x:0, y:0, z:0)
Point Cloud - a collection of point in space, generally representing a multitude of points on a physical object.

Basics

Point clouds can have color, but they can't represent textures

Basics

In order to have textures, a point cloud but be converted to a mesh.

There are many algorithms to convert clouds to meshes.
It is difficult for a single algorithm to perfectly reconstruct any arbitrarily shaped point cloud.
This a practical sense, this is the hardest (experimental) part of the process.

Basics

Structure from motion (SfM)

A range imaging technique; it refers to the process of estimating three-dimensional structures from two-dimensional image sequences which may be coupled with local motion signals.