big geospatial data + deep learning + high performance ...€¦ · big geospatial data + deep...
TRANSCRIPT
GXP Xplorer and SOCET GXP are registered trademarks of BAE Systems. All other brands, product names, and trademarks are property of their respective owners.
This document gives only a general description of the product(s) or service(s) offered by BAE Systems and, except where expressly provided otherwise, shall not
form part of any contract. From time to time, changes may be made in the products or conditions of supply. Approved for public release on 02/25/2016.
ES-GEO-22316-0004 1
GEOSPATIAL EXPLOITATION PRODUCTS
Big Geospatial Data + Deep Learning + High Performance
Computing = Geospatial Intelligence
Bingcai Zhang
Tech Fellow
ES-GEO-22316-0004
Do We Have Enough Parking?
– Demo
– 300 drone images
– GSD = 3.5cm
– 1600 cars detected
– 99% detection accuracy
– 6 pixel positional accuracy
– 10 degree orientation accuracy
– 0.001% false positive error rate
– Data from Prof. Dunn
– Model trained with 7.5cm GSD
2
ES-GEO-22316-0004
What Is Deep Learning?
It works just like the brain (least favorite definition according to LeCun)
3
car
0.5
not car
0.5
car
1.0
not car
0.0
IEEE SPECTRUM
ES-GEO-22316-0004
Simplicity Learning
– Object detection very complex
– Breakup objects
• Learn one type of object at a time
• Detect one type of object at a time
– Inspired by 4 year old pre-school best learning practice
• Learn one alphabet per week
• Learn letter “A” five days in a row (reinforcement learning)
– Inspired by drug discovery
• One specific drug for one specific disease
• No panacea
– Based on two decades of research experience
• Transform a complex problem into its simplest components
• Solve each component one at a time
4
ES-GEO-22316-0004
Data Normalization vs. Data Augmentation
– Scale normalization vs. scale augmentation
– Color normalization vs. color augmentation
– Rotation normalization (geospatial images)
5
ES-GEO-22316-0004
Simplicity Learning vs. Non-Simplicity Learning
6
vehicle and anything else vehicle, stop sign, and
anything else
ES-GEO-22316-0004
Handcrafted Automatic Feature Extraction
7
ES-GEO-22316-0004
3D Features (3D Glasses)
8
ES-GEO-22316-0004
Rotation Variant Object Detection
9
-30
-20
-10
0
10
20
30
-88 -80 -72 -64 -56 -48 -40 -32 -24 -16 -8 0 8 16 24 32 40 48 56 64 72 80 88
avg
min
max
std
ES-GEO-22316-0004
Rotation Variant Object Detection… 2
10
0
0.2
0.4
0.6
0.8
1
1.2
-88 -80 -72 -64 -56 -48 -40 -32 -24 -16 -8 0 8 16 24 32 40 48 56 64 72 80 88
probability
probability
ES-GEO-22316-0004
Rotation Variant Object Detection… 3
11
ES-GEO-22316-0004
Rotation Variant Object Detection… 4
12
ES-GEO-22316-0004
Singular Classification
13
c
j
jz
iz
e
esoft
1
:max
Open-ended negative training examples problem
Not work just like the brain
Three new algorithms reducing false positive by 10
times
ES-GEO-22316-0004
Human vs. Machine
14
– Human: Radoslav Gaidadjiev
• Two master’s degree
• Twenty years of experience with imagery
– Machine: DeepObject with one K40 GPU
– Human achieved accuracy of 99.9%
• Understand parking lot
– Machine achieved accuracy of 99.3%
– Demo
• Everyone participates human vs. machine
ES-GEO-22316-0004
Quality Training Samples = $
– Quality training samples are the new currency in deep learning
• Non-deep learning:
• AOD failed to recognize a car and an image analyst found this missing car
• A DR is generated and the cost to fix this DR is $1000
• Deep learning:
• This missing car could be automatically collected as a positive training example and
added to the training sample database
• Train deep learning network again with the new training example database
• Cost could be as low as $10
– Potential cost saving is very significant
15
ES-GEO-22316-0004
Mistakes = Quality Training Samples
– We learn from our mistakes (so does deep learning)
• Not all training samples are created equal
• Mistakes are more likely to have greater gradient
• Have stronger influence on decision boundary
• Quality of training samples is as important as quantity of training samples
• Data augmentation increases quantity
• Mistakes increases quality
– Future geospatial intelligence software should collect users intelligence
– Every mistake could translate into an enhancement to geospatial
intelligence software
16
ES-GEO-22316-0004
Future Intelligent Geospatial Intelligence System
– An intelligent system that could become smarter and smarter by learning
from its mistakes
– An intelligent system that could detect and monitor defense relevant
objects at 99% accuracy
– With 99% accuracy, this may be the game changer in geospatial
intelligence domain
– Significantly reduce software engineering and enhancement costs
17
ES-GEO-22316-0004
Questions?
Dr. Bingcai Zhang
858-592-5218
18