Lecture_14_object_detection_segmentation Project Proposal Instructions
User Manual:
Open the PDF directly: View PDF 
.
Page Count: 8
•Can work alone or in a group (up to 4 people), required effort will scale with # of people
•Select a “base” dataset (online, or from a list I’ll make)
•Simulate parameters of a physical (imaging) system with base dataset
•Train deep neural net with simulated dataset
•Report results
Class project details
Class project details
What you’ll need to submit:
1) The project’s source code
2) A short research-style paper (3 pages minimum, 5 pages maximum) that includes an
introduction, results, a discussion section, references and at least 2 figures
3) A completed web template containing the main results from the research paper
4) An 8-minute presentation that each student will deliver to the class
•Can work alone or in a group (up to 4 people), required effort will scale with # of people
•Select a “base” dataset (online, or from a list I’ll make)
•Simulate parameters of a physical (imaging) system with base dataset
•Train deep neural net with simulated dataset
•Report results
Example project topics:
Can we design a new lens/transducer/antenna shape to improve classification of X?
What is the tradeoff between image resolution and classification accuracy for X?
Can we determine an optimal set of colors to improve fluorophore distinguishability?
If we capture 2 images that are overlapped on one sensor, what is the best way to pre-blur them to then be able to tell them
apart? Or to be able to classify them together?
If we just had a few sensors, how should be arrange them e.g. a mask to be able to predict the position of X?
Is there some optimal shift-variant blur that we can to use for a particular task?
Or, given a shift-variant PSF image, can we establish a good deconvolution using locally connected layers?
What is the optimal way to layout filters on a sensor to capture a color image for classification? Or an HDR image?
HDR image generation with filters over pixels –what is optimal design?
What if we could make a sensor with different sized pixels –how should they be laid out to achieve the best X?
Class project –what are the first steps?
1. Think about it!
2. Discuss with your friends/others in the class (feel free to use Slack!)
3. Schedule a short 15 meeting with me:
•Friday 3/1, 3:30pm – 6:00pm
•Next Monday 3/4, 1:00pm – 3:30pm
•Next Wednesday 3/6, 10:00am – 1:00pm
4. Start to write-up a proposal
•General aim: 1 paragraph with specification of physical layer
•Discussion: (a) data source(s), (b) expected simulations, (c) expected CNN, (d)
quantitative analysis of physical layer (comparison, plot, etc).
•Project proposal due date: Thursday March 7, 2019
•Revised project proposal due date: Tu es day Mar ch 19, 2019
Example project topics:
Can we design a new lens/transducer/antenna shape to improve classification of X?
What is the tradeoff between image resolution and accuracy for X (classification, segmentation, etc.)? What if we had access to n
low-resolution cameras –how might we position them to get the best performance?
Can we determine an optimal set of colors to improve fluorophore distinguishability?
If we capture 2 images that are overlapped on one sensor, what is the best way to pre-blur them to then be able to tell them
apart? Or to be able to classify them together?
If we just had a few sensors, how should be arrange them e.g. a mask to be able to predict the position of X?
Is there some optimal shift-variant blur that we can to use for a particular task?
Or, given a shift-variant PSF image, can we establish a good deconvolution using locally connected layers?
What is the optimal way to layout filters on a sensor to capture a color image for classification? Or an HDR image?
HDR image generation with filters over pixels –what is optimal design?
What if we could make a sensor with different sized pixels –how should they be laid out to achieve the best X?
What is the tradeoff between image resolution and accuracy for image segmentation? What if we had
access to n low-resolution cameras –how might we position them to get the best performance?
I"propose"to"simulate"the"classification"performance"of"a"new"
type"of"microscope," which"will"have"3"different"lenses"and"
sensors."Each"lens"and"sensor"will"capture"an"image"of"a"flat"
object"from"a"unique" angular" perspective,"and"the"image"
classification"will"be"performed" with"all"of"the"data."The"physical"
parameter"that"I"will"optimize" is"the"angle"of" tilt"of"each"lens"
with"respect"to"the"object"to"maximize"classification"accuracy.
What is the tradeoff between image resolution and accuracy for image segmentation? What if we had
access to n low-resolution cameras –how might we position them to get the best performance?
I"propose"to"simulate"the"classification"performance"of"a"new"
type"of"microscope," which"will"have"3"different"lenses"and"
sensors."Each"lens"and"sensor"will"capture"an"image"of"a"flat"
object"from"a"unique" angular" perspective,"and"the"image"
classification"will"be"performed" with"all"of"the"data."The"physical"
parameter"that"I"will"optimize" is"the"angle"of" tilt"of"each"lens"
with"respect"to"the"object"to"maximize"classification"accuracy.
a2
a1a3
Optimize:"a1-3"to"max."classification"
accuracy"of"imaging"system"
Image 1
Image 2 Image 3
Classification"CNN
What is the tradeoff between image resolution and accuracy for image segmentation? What if we had
access to n low-resolution cameras –how might we position them to get the best performance?
I"propose"to"simulate"the"classification"performance"of"a"new"
type"of"microscope," which"will"have"3"different"lenses"and"
sensors."Each"lens"and"sensor"will"capture"an"image"of"a"flat"
object"from"a"unique" angular" perspective,"and"the"image"
classification"will"be"performed" with"all"of"the"data."The"physical"
parameter"that"I"will"optimize" is"the"angle"of" tilt"of"each"lens"
with"respect"to"the"object"to"maximize"classification"accuracy.
a2
a1a3
Optimize:"a1-3"to"max."classification"
accuracy"of"imaging"system"
Image 1
Image 2 Image 3
Classification"CNN Dataset:"12,500"images"of"4"types"of"blood"cell
https://www.kaggle.com/paultimothymooney/blood-cells
Simulation:"Treat"each"image"as"a"thin"2D"object"and"is"coherently"illuminated."
Assume"each"camera"captures"a"unique"component"of"the"object"spectrum,"which"
will"vary"as"a"function" of"a1-3."Start"by"neglecting"size"and"shape"of"each"camera.
CNN:"Digital"layer:"Alexnet."Physical"layer:"simulate"object"spectrum,"sample"object"
spectrum"re-centered"by"angle"a1-3 (which"are"weight"variables),"form"images"and"
classify"them"together
Quantitative"analysis:"I"will"plot"classification"performance"as"a"function"of"the"
number"of"allowed"cameras"for"a"fixed"CNN"architecture,"and"will"also"compare"the"
classification"performance"to"the"case"of"a"single"image"
