Pix4D Training

July 2013
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: Calibration
Step2: Point densification
Step3: DSM and orthomosaic

Output files usage

Software quick guide

Introduction
Pix4D converts thousands of images into
3D Point Cloud
3D Digital Surface Model
Orthomosaic
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: calibration
Step2: point densification
Step3: DSM and orthomosaic

Output files usage

Software quick guide

From 2D to 3D
From 3D to 2D From 2D to 3D

One camera Two images

Far appears smaller Triangulate to get
than near on image sense of depth
 From 2D to 3D: Stereoscopy
3D point

2D Keypoint
2D Keypoint

Camera model Camera model

(focal length, (focal length,
principal point, principal point,
lense distortion) Camera position lense distortion)
and rotation
Camera calibration
Exact camera model and position needed for
reconstruction

GPS tag not precise enough (1-10 meters)

Pixels contain geometrical information about

the exact camera position (1-10mm)

Calibration is about computing exact

camera position/rotation/model from pixels
alone
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: calibration
Step2: point densification
Step3: DSM and orthomosaic

Output files usage

Software quick guide

 3 main processing steps

Step 1: initial Step 2: point Step 3: DSM and

processing densification orthomosaic
Images Calibrated cameras Point clouds
=> calibrate cameras => point clouds => DSM and orthomosaic

n ing tim n ing

c tio c h o p G CP atio t e r a tio
n
a ge
s
d ing
tra t l / f i c f i l r n
t ex s ma od
e
G PS e ns
i
int g e ne ic im c ble
in t m n d o a i
po p oin e ra a tio int
P
SM m os os
a
y y m c o D o m
Ke Ke Ca e olo P
Orth rth
o
G O
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: calibration
Step2: point densification
Step3: DSM and orthomosaic

Output files usage

Software quick guide

 Camera calibration
Step 1: initial processing
Images => calibrate cameras

tio
n
hing ptim CP
trac atc elo
PS
/G
ex sm mo
d
in t int onG
y po ypo era ati
Ke Ke Cam oloc
Ge

600'000
Example: 60'000 keypoints Average of 6'000 matches => Average of GPS
per image matches per pair 150'000 3D OR
of images Points for 100 Min 3 GCP
images project
Camera calibration
1. Extract keypoints
a. interesting areas, high contrast
b. create "signature" based on surrounding image content
2. Match keypoints
a. Find neighboring images using GPS
b. Match all keypoints between these images
3. Camera model
a. Fit a camera model to the matched keypoints
b. Triangulate matches to make 3D points
c. Optimize camera model and keypoint matches
d. Add more images (and keypoints), and repeat
4. Geolocation
a. Use GPS tags average for approximate geolocation
b. Use GCP for precise geolocation
Extract Keypoints
Good image:
rocks,bushes,dirt
buildings, urban
more than 10MP
extract > 10'000 keypoints

Difficult images:
sand, snow, fog
blurry, out of focus
Overexposed, underexposed
less than 3MP
extract < 100 keypoints
Match keypoints
Easy to match
images with high # keypoints
images with high overlap

Hard to match
images with low # keypoints
trees at low altitude
extreme angle with no transition
low overlap

Impossible to match
Reflective surface like water
Moving objects
Getting enough matches
Enough matches:
# matches > 1000 per image pair
Depends on overlap, image size and visual content
If small image size => more overlap required
If low visual content => more overlap required
Images of 12 MP => 75% overlap recommended in most cases

>1000 matches <100 matches

75% overlap 20% overlap

enough matches not enough matches
Flight plan
Ideal flight plan
regular grid flight plan
easy terrain: 75% frontal, 50% side
difficult terrain: 85% frontal, 60% side

Difficult flight plan

low overlap
multiple images at same location
corridor mapping
high difference in altitude (>2xGSD)
Camera Calibration: quality report

Ensures quality of project

Issues with project
To solve 99% of issues: Increase overlap or fly again
Proposed workflow
Camera calibration: accuracy
Good calibration

Error in calibration: error propagation

Sufficient matches prevent calibration

errors
(> 1000 matches/image)
Georeference
Project size, orientation and position unknown
GPS tags
approximate position,orientation,size
good synchronisation, high number of GPS tags => better approximation
~20 meter shift in random direction

GCP
If GCP, GPS tags are not used
Minimum 3 GCPs
Results in same coordinate system as GCPs
Geolocation: accuracy
GCPs
well distributed in dataset, measured below GSD
5-8 usually enough for 1000 images
Quality report shows reprojection error

Good calibration + GCP: error at most 1-2 times

the GSD

Verification points
Error is minimized at GCPs: not optimal to assess
global accuracy
Verification points are not used during optimization
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: calibration
Step2: point densification
Step3: DSM and orthomosaic

Output files usage

Software quick guide

Point densification
Step 2: point densification
Calibrated cameras => cloud of points

n g
atio t erin
c fil
nsifi int
t de Po
i n
Po

Example: 8 millions 3D 2 millions filtered

points and triangulated
3D points
Point densification
Visual content
3D points computed where there is visual content
There may be "holes" in point cloud
No visual, no 3D (example: no ground under trees)
Roads, rooftops, may contain little visual content
Less visual => less accuracy/less points

Scale
Trees visual content only at lower resolution
Use "high tolerance" to model trees
Expect lower accuracy at trees
Point densification: filtering
Noise filtering
improves quality
2.5D only

Smoothing
removes redundant points
makes point cloud easier to manage (smaller files generated)
2.5D only
Point densification: accuracy
Point cloud accuracy
Good camera: error at most 1-2 times the GSD
up to 3 times the GSD at areas with low visual content

Global Accuracy
Calibration+Georeference+Point Cloud
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: calibration
Step2: point densification
Step3: DSM and orthomosaic

Output files usage

Software quick guide

DSM and Orthomosaic
Step 3: DSM and orthomosaic
Point cloud
=> DSM and orthomosaic

tion es ing
era ag d
n i m blen
ge saic aic
M mo s
DS tho o mo
Or th
Or
DSM generation

Point cloud => Regular grid of elevation (2.5D)

Interpolation where no points => no holes

DSM accuracy
Interpolation at place with no visual information

Global Accuracy
Calibration+Georeference+Point
Cloud+Interpolation
Photo Stitching VS
Orthorectification
Glue images together Undistort perspective using the
3D model and blend images

Works only if terrain perfectly flat Handles any terrain type

Error accumulation, only small Handles large dataset
dataset Supports Georeference perfectly
No good georeference support Preserves distances and
without GCP becomes measurable
Most distances not preserved Requires high number of
Requires low number of matches/keypoints (>1000) to
matches/keypoints (<100) generate the 3D model
Orthorectification
Orthorectify image
1) calibrate camera
UAV image 2) project image on 3D model
with 3) generate ortho image by looking
perspective from above

UAV image with perspective Computed 3D model Undistorted orthoimage

Facade Same as satellite
Roof not correct size Roof correct size
Orthorectification artefacts
3D model not always perfect
Edges of buildings
Small details
trees (branches smaller than pixels)
lamppost, fences

Orthorectification depends on surface model =>

artifacts at 3D model errors
Original image tilted: area not
visible producing artefacts

3D model not as precise as

texture, leading to "wobbly"
edges
Blending and Editing

Default: automated blending

Scene editor:
choose image
correct brightness
selective use of orthorectification to generate
visually pleasing mosaics
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: calibration
Step2: point densification
Step3: DSM and orthomosaic

Output files usage

Software quick guide

 Generated files
Quality report
Camera position & parameters
Raw point cloud
(ply,las)

Filtered point cloud (ply/las)

regular point cloud
DSM in tiff
individual ortho and
planar
orthomosaic in
tif, KML tiles
Step 1: initial Step 2: point Step 3: DSM and
processing densification orthomosaic

n ing tim n ing

c tio c h o p G CP atio t e r a tio
n
a ge
s
d ing
tra t l / f i c f i l r n
t ex s ma od
e
G PS e ns
i
int g e ne ic im c ble
in t m n d o a i
po p oin e ra a tio int
P
SM m os os
a
y y m c o D o m
Ke Ke Ca e olo P
Orth rth
o
G O
 Generated files
Pix4uav outputs Format Possible use Examples of compatible
software

Raster orthomosaic geoTIFF (.tiff) Area overview ArcGIS

KML tiles (.png/.kml) Digitize buildings Global Mapper
Annotate areas QuantumGIS
Overlay in GIS package
AutoCAD
Analyze spectral bands
Google Earth

Undistorted images TIFF (.tiff) Stereo Viewing ImageMaster

Individual ortho/planar geoTIFF (.tiff) Seamline editing OrthoVista

3D .las, .ply, .ascii Visualization ArcGIS

point cloud Surface Editing Global Mapper
DSM generation AutoCAD
DTM generation
Quick Terrain Reader
3D Reshaper
Trimble RealWorks Viewer

Raster digital surface geoTIFF (.tiff) Analyze surface ArcGIS

model (DSM) Measure volumes Global Mapper
Generate contour lines QuantumGIS
Draw breaklines Quick Terrain Reader

3D mesh with texture Wavefront (.obj) Render in animation package AutoCAD

Visualize small projects Bentley Pointools View
CC Viewer
3D Reshaper
Use of generated files

3D point cloud Raster DSM 3D Mesh

Editing best (VRMesh) possible possible

Analyzing mainly transform to raster best (ArcGIS/Global small projects only

DSM Mapper)

Visualization good (Quick Terrain good (ArcGIS/Global small projects only

Reader) Mapper)
Plan
Intro

Stereoscopy/calibration theory

Processing steps
Step1: calibration
Step2: point densification
Step3: DSM and orthomosaic

Output files usage

Software Demo
Software Demo: create project

3 Step wizard
1&2)
- create project, import images
- is also used to "merge" overlapping
project. Split and merge when computer
resources not sufficient, or when
calibration issues between areas of
project.

3)
geotags and camera model selection
- automatic geotag, or using text file
- automatic camera model selection,
or using database
- at least three valid geotag required
 Software Demo: main interface

1) Processing
button

2) Start button
Steps in red if not yet
processed. Checkbox
allows run of individual
step.

Initial Processing: Point densification: Orthomosaic/DSM:

Option: rapid,full Option: density Option: resolution

127 images rapid: 3 min, low optimal: 28 min auto will select average
12 MP Canon Ixus accuracy Choose high computed GSD: 15 min
quadcore laptop full: 14 minutes tolerance if forest Can be replaced by any
other number
 Software Demo: GCP editor
1) Add GCP/control/verification
using file OR right click

3) Measure GCP in at least 2

images per GCP, 5
recommended

5) If initial processing is
already done, bundle
2) List of images 4) Import/export with GCP allows for
use top icons to sort clicked points without faster re-processing
closest images coordinates
use right click to remove
measurements
Software Demo: Scene Editor
4) Change/Add cells using 5) Blend mosaic
polygon/line/points tools

2) Select ortho/planar
to improve visual
aspect

3) Choose most
appropriate image in
the list (use keyboard
up/down arrow).
Good to remove
blurry images/ moving
cars

1) Select a cell by
clicking
Thank you