Explore a keynote address from the Robust Vision Challenge 2020 focusing on robustness across the data abundance spectrum. Delve into the challenges of modern computer vision, including data-efficient learning and benchmarking for large-scale datasets. Examine the development of LVIS (Large Vocabulary Instance Segmentation) and its innovative approach to federated dataset design. Learn about the difficulties in annotating thousands of categories and discover data-driven category discovery techniques. Gain insights into best practices for handling noise in datasets and get information about the LVIS Challenge at ECCV 2020.
Keynote: Ross Girshick - Robustness Across the Data Abundance Spectrum
Overview
Syllabus
Einleitung
Axes of Robustness
The Numbers You See in Papers
The Reality
Back to the Drawing Board
It All Starts with Naming Things
Naming is for Communication
Human Human Communication
Object Spotting Game car
Task: Detect Everything!
Problem 2: Data Efficient Learning
Problem 1: Benchmarking
Building LVIS
PASCAL VOC and COCO were "easy" to Build
Troubles with 1000's of Categories
Facing These Problems, What Can We Do?
Federated Dataset Design to the Rescue
Annotation Pipeline
LVIS Annotations Quality
Where does the vocabulary come from?
Data Driven Category Discovery
The Long Tail is Inescapable! 100
Results on LVIS v1 validation set
Best Practices - Know What's Noise
The LVIS Challenge at ECCV 2020!
Taught by
Andreas Geiger
