Budget-aware Semi-Supervised Semantic and Instance Segmentation - - PowerPoint PPT Presentation

Budget-aware Semi-Supervised Semantic and Instance Segmentation

Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Women In Computer Vision - CVPR 2019

Motivation

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Semantic segmentation Instance segmentation Pixel-level annotations are expensive!

Motivation

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Semantic segmentation Instance segmentation

Solution: Weakly and Semi-supervised methods!

Motivation

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Semantic segmentation Instance segmentation

Solution: Weakly and Semi-supervised methods!

Contributions

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

1. We unify the segmentation benchmarks regardless the training setting and the supervision signals comparing them in terms of the total annotation cost.

Contributions

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

1. We unify the segmentation benchmarks regardless the training setting and the supervision signals comparing them in terms of the total annotation cost. 2. We experiment with a semi-supervised pipeline and test it in Pascal VOC for semantic and instance segmentation, outperforming previous works at low annotated budgets.

Contributions

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

1. We unify the segmentation benchmarks regardless the training setting and the supervision signals comparing them in terms of the total annotation cost. 2. We experiment with a semi-supervised pipeline and test it in Pascal VOC for semantic and instance segmentation, outperforming previous works at low annotated budgets. 3. We show that for low annotation budgets, it’s more convenient having fewer but stronger-labeled data

• ver having larger weakly-annotated sets.

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Semi-Supervised Pipeline Semi-Supervised Pipeline

Annotation network

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Semi-Supervised Pipeline Semi-Supervised Pipeline

Annotation network Segmentation network

Experimental validation for Pascal VOC

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Semantic segmentation

Annotation cost Segmentation quality

We used DeepLab v3+ for both the annotation and segmentation network

Experimental validation for Pascal VOC

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Instance segmentation

Annotation cost Segmentation quality

We used RSIS for both the annotation and segmentation network

Comparison to other works

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Semantic segmentation Instance segmentation

Annotation cost Segmentation quality

Visualization for semantic segmentation

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Visualization for instance segmentation

Budget-aware Semi-Supervised Semantic and Instance Segmentation: Miriam Bellver, Amaia Salvador, Jordi Torres, Xavier Giro-i-Nieto

Thanks for your attention!