GenSeg

Instance Segmentation- MP

Purpose

Common Techniques - intr Star Dist

StarDist - MP

Regression

Network Architecture

Loss

Disadv

genseg proposal -ML

general ideas - any shape can be modeled as a collection of gaussians

memory advantage

some other advantages ( we will talk abt in next slides)

BBlocks of StarDistML

1. Estimation of Distance parameters from Masks

2. Training

3. Non Maxima + Inference

 

BBlocks 1 of GensegMP

1. IOU loss

2. Equations ( c and t)

3. predicts 6 parameters

4. Negative, Unormalized GMM

5. UpperBound - best we can get

Results of GensegMP

3 figures, IOU, StarDist (64, 128, 256)

1. Elephant

2. Potato

3. Fingers

(table - 3 rows, 4 cols)

BBlock 2 - Training ML

1. re-explain stardist loss, architecture drawing

2. explain our strategy

3. talk about distance of gaussians (our loss), also KL divergence

4. our t and c are know, so we only predict gaussian parameters and compare GMMs

 

 

 

BBlock 3 - Inference ML

1. Stardist NMS

2. Our NMS (L2 loss, IOU loss)

 

 

Timeline

1. Datasets  1 and 2

2. mention some steps

 

 

BBlock 2 - Training

1. re-explain stardist loss, architecture drawing

2. explain our strategy

3. talk about distance of gaussians (our loss)




Overview

  • StarDist
  • Estimation of Shape Description
  • Results
  • Future Work
  • Timeline