Congealing or Finding the Platonic Gate Jason Fennell & Joe - - PowerPoint PPT Presentation

congealing or finding the platonic gate
SMART_READER_LITE
LIVE PREVIEW

Congealing or Finding the Platonic Gate Jason Fennell & Joe - - PowerPoint PPT Presentation

Congealing or Finding the Platonic Gate Jason Fennell & Joe Simons Outline Sketching as a whole Gate level recognition Our solution: Congealing Motivation Algorithm Results Future work Sketching vs. Design


slide-1
SLIDE 1

Congealing

  • r

Finding the Platonic Gate

Jason Fennell & Joe Simons

slide-2
SLIDE 2

Outline

 Sketching as a whole

 Gate level recognition

 Our solution: Congealing

 Motivation  Algorithm  Results

 Future work

slide-3
SLIDE 3

Sketching vs. Design

 Sketching is fast and intuitive  But there is no automated verification or simulation for

sketches

 Putting a sketch into a design program is tedious  Solution: Make a computer do it!

slide-4
SLIDE 4

But Sketch Recognition is HARD!

slide-5
SLIDE 5

Sketch Recognition Subtasks

Stroke Fragmentation Stroke Grouping Symbol Recognition

NAND

slide-6
SLIDE 6

Sketch Recognition Subtasks

Stroke Fragmentation Stroke Grouping Symbol Recognition

NAND

slide-7
SLIDE 7

Congealing

 Based on Learning from One Example Through Shared

Densities on Transforms

 Individual instances of gates vary widely  We want to create a “platonic gate” to do

recognition against

slide-8
SLIDE 8

Congealing

a a a a a a a a

The platonic a Instances of a

slide-9
SLIDE 9

Congealing

 Want a computer to be able to do find platonic images

automatically

 Assume that there is some transformation from the

platonic image to any gate you draw

 Given a set of gates, find the inverse of these

transformations

 We assume transforms are affine

 Scale  Shift  Shear  Rotate

slide-10
SLIDE 10

Training

 Want to minimize the summed pixel-wise entropy  Average image  Metric

 Binary entropy function

slide-11
SLIDE 11

Training

  • 1. Apply affine transform to an individual image
  • 2. If the transformation decreases the total entropy,

keep it

  • 3. Repeat 1 and 2 for each image and possible affine

transform

  • 4. Repeat 1-3 until improvement stops
slide-12
SLIDE 12

Training Results

slide-13
SLIDE 13

Training Results

slide-14
SLIDE 14

Classification

 Congeal candidate image against a sequence of

average images

 This creates a version of the candidate image in

platonic space

 Use a simple distance-based classifier on images in

platonic space.

slide-15
SLIDE 15

Future Work

 Create a generic function that maps from affine

transforms to total entropy  Will allow us to use a variety of numerical methods

 Testing of the classifier with several metrics  Recognition of sub-parts of gates to aid in

grouping stage of sketch recognition

slide-16
SLIDE 16

Questions?