Explore the challenges and solutions in computer vision models through this comprehensive tutorial from CVPR'22. Delve into standard visual recognition pipelines, benchmark performances, and dataset biases. Examine adversarial examples, benchmark challenges, and domain adaptation techniques. Learn about adapting to imbalanced data, self-training methods, and the SENTRY approach for selective entropy optimization. Investigate performance degradation from bias, geographic diversity in data, and the concept of responsible vision. Gain insights into making object recognition work effectively for everyone across different geographical contexts.
Overview
Syllabus
Intro
Standard Visual Recognition Pipeline
Visual Recognition Benchmark
Benchmark Performance
Dataset Bias
Adversarial Examples
Benchmark Challenge Adversar
RobustNav Dynamics Corruptid
Domain Adaptation: Train on Source Test on
Domain Adversarial Adaptatio
Adapting to Imbalanced Data
Adaptation with Self-Training Entropy Minimization for UDA
SENTRY: Selective Entropy Optimization Selective Entropy Minimization
Selective Entropy Loss
SENTRY Results: Image Classification
SENTRY Results: MiniDomainNet
Extension to Semantic Segmentation
Performance Degradation from Bias
Geographic Bias
Does object recognition work for everyone?
Can domain adaptation make obj rec work for everyone?
Geographically diverse data
Additional challenges in GeoDA
Summary: Responsible Vision
Taught by
Bolei Zhou
