Detecting dangers in traffic scenes
Event details
| Date | 26.06.2018 |
| Hour | 13:30 › 15:30 |
| Speaker | Krzysztof Lis |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Dr. Martin Rajman
Thesis advisor: Prof. Pascal Fua
Thesis co-advisor: Dr. Mathieu Salzmann
Co-examiner: Dr. Micahël Thémans
Abstract
Self-driving cars will have to flexibly react to unpredictable dangers.
However many modern computer vision techniques rely on learning from datasets containing only typical situations.
Our goal is to detect unusual dangers (animals, rare vehicles and obstacles, anything not present in the datasets) in images of traffic scenes.
We discuss three articles related to our task:
an instance-aware semantic segmentation method InstanceCut - the base model of our scene understanding;
a class-agnostic object proposal generator DeepMask, capable of generalizing to previously unseen object categories;
and a conditional GAN called pix2pixHD which generates images from semantic labels, possibly allowing us to detect errors in semantic segmentation in an unsupervised way.
We present our results and plans for further research.
Background papers
InstanceCut: from Edges to Instances with MultiCut, by Kirillov, A., et al.
Learning to Segment Object Candidates, by Pinheiro, P., et al.
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs, by Wang, t-C., et al.
Exam president: Dr. Martin Rajman
Thesis advisor: Prof. Pascal Fua
Thesis co-advisor: Dr. Mathieu Salzmann
Co-examiner: Dr. Micahël Thémans
Abstract
Self-driving cars will have to flexibly react to unpredictable dangers.
However many modern computer vision techniques rely on learning from datasets containing only typical situations.
Our goal is to detect unusual dangers (animals, rare vehicles and obstacles, anything not present in the datasets) in images of traffic scenes.
We discuss three articles related to our task:
an instance-aware semantic segmentation method InstanceCut - the base model of our scene understanding;
a class-agnostic object proposal generator DeepMask, capable of generalizing to previously unseen object categories;
and a conditional GAN called pix2pixHD which generates images from semantic labels, possibly allowing us to detect errors in semantic segmentation in an unsupervised way.
We present our results and plans for further research.
Background papers
InstanceCut: from Edges to Instances with MultiCut, by Kirillov, A., et al.
Learning to Segment Object Candidates, by Pinheiro, P., et al.
High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs, by Wang, t-C., et al.
Practical information
- General public
- Free
Contact
- EDIC - [email protected]