際際滷

際際滷Share a Scribd company logo

Measuring the impact of label noise on semantic segmentation using rastervision

L
Louisa Diggs

Quantifying Error in Training Data for Mapping and Monitoring the Earth System - A Workshop on Quantifying Error in Training Data for Mapping and Monitoring the Earth System was held on January 8-9, 2019 at Clark University, with support from Omidyar Networks Property Rights Initiative, now PlaceFund.

1 of 19
Download to read offline
Measuring the impact of label noise on semantic segmentation using Lewis Fishgold
Outline Motivation: Using noisy OSM labels for building segmentation Prior experiments on effects of label noise Our experiments on building segmentation Implementation using Raster Vision
Motivation: Building Segmentation using Noisy OSM Labels Training chip with small label errors Predictions over Georgetown, Guyana
Prior Work From Deep Learning is Robust to Massive Label Noise, Rolnick et al. 2018
Prior Work From Deep Learning is Robust to Massive Label Noise, Rolnick et al. 2018
Our Experiments How does noise affect accuracy for our use case? Use Spacenet Vegas Buildings dataset Semantic segmentation Satellite Imagery More realistic noise Deleting labels Shifting labels Warning: very preliminary work
Randomly Deleted Labels
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
Randomly Shifted Labels
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
Measuring the impact of label noise on semantic segmentation using rastervision
References Deep Learning is Robust to Massive Labeling Noise: https://arxiv.org/abs/1705.10694 Spacenet Dataset: https://spacenetchallenge.github.io/ Raster Vision: https://github.com/azavea/raster-vision Code for these experiments: https://github.com/azavea/raster-vision-experiments/pull/1

More Related Content

Measuring the impact of label noise on semantic segmentation using rastervision

  • 1. Measuring the impact of label noise on semantic segmentation using Lewis Fishgold
  • 2. Outline Motivation: Using noisy OSM labels for building segmentation Prior experiments on effects of label noise Our experiments on building segmentation Implementation using Raster Vision
  • 3. Motivation: Building Segmentation using Noisy OSM Labels Training chip with small label errors Predictions over Georgetown, Guyana
  • 4. Prior Work From Deep Learning is Robust to Massive Label Noise, Rolnick et al. 2018
  • 5. Prior Work From Deep Learning is Robust to Massive Label Noise, Rolnick et al. 2018
  • 6. Our Experiments How does noise affect accuracy for our use case? Use Spacenet Vegas Buildings dataset Semantic segmentation Satellite Imagery More realistic noise Deleting labels Shifting labels Warning: very preliminary work
  • 19. References Deep Learning is Robust to Massive Labeling Noise: https://arxiv.org/abs/1705.10694 Spacenet Dataset: https://spacenetchallenge.github.io/ Raster Vision: https://github.com/azavea/raster-vision Code for these experiments: https://github.com/azavea/raster-vision-experiments/pull/1