learning specific-class segmentation from diverse data

13
Learning Specific-Class Segmentation from Diverse Data M. Pawan Kumar, Haitherm Turki, Dan Preston and Daphne Koller at ICCV 2011 VGG reading group, 29 Nov 2011, presented by Varun Gulshan

Upload: sanura

Post on 22-Feb-2016

47 views

Category:

Documents


0 download

DESCRIPTION

Learning Specific-Class Segmentation from Diverse Data. M. Pawan Kuma r, Haitherm Turki , Dan Preston and Daphne Koller at ICCV 2011. VGG reading group, 29 Nov 2011, presented by Varun Gulshan. Semantic image segmentation. Main idea. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Learning Specific-Class Segmentation from Diverse Data

Learning Specific-Class Segmentation from Diverse Data

M. Pawan Kumar, Haitherm Turki, Dan Preston and Daphne Koller at ICCV 2011

VGG reading group, 29 Nov 2011, presented by Varun Gulshan

Page 2: Learning Specific-Class Segmentation from Diverse Data

Semantic image segmentation

Page 3: Learning Specific-Class Segmentation from Diverse Data

Main idea

• High level: Getting fully labelled data for training is expensive, use other easily available ‘diverse’ data for learning (bounding boxes, classification labels for image).

Tags: Car, peoplePerson bounding box

Page 4: Learning Specific-Class Segmentation from Diverse Data

Implementing the idea

• The bounding box/image classification data is incomplete for segmentation, fill in the missing information using latent variables.

• Setup the training cost function using latent variables. Use their self-paced learning algorithm for Latent-SVM’s [NIPS2010] to optimise the training cost function.

• While inferring latent variables, make sure latent variable estimation is consistent with the weak annotation. Setting up the inference problems to ensure this condition.

Page 5: Learning Specific-Class Segmentation from Diverse Data

Energy function without latent variables

Notation:

Image

Parameters to be trained

Joint feature vector (essentially the terms of a CRF)

Page 6: Learning Specific-Class Segmentation from Diverse Data

Structured output training

Ground truth labels

Loss function

Page 7: Learning Specific-Class Segmentation from Diverse Data

Introducing latent variables

Page 8: Learning Specific-Class Segmentation from Diverse Data

Introducing latent variables

But we don’t know what hk is (its latent), so maximise it out.

Page 9: Learning Specific-Class Segmentation from Diverse Data

Introducing latent variables

Page 10: Learning Specific-Class Segmentation from Diverse Data

Self-paced optimisation

Page 11: Learning Specific-Class Segmentation from Diverse Data

Self-paced optimisation

Indicator variable to switch off the harder cases.

Page 12: Learning Specific-Class Segmentation from Diverse Data

Second idea: Latent variable estimation

The algorithm involves estimating annotation consistent latent variables in the following equation:

More precisely

Page 13: Learning Specific-Class Segmentation from Diverse Data

Move to white-board

Me

You

Beware of Equations