Rethinking Class-Balanced Methods for Long-tailed Visual Recognition - PowerPoint PPT Presentation

Rethinking Class-Balanced Methods for Long-tailed Visual Recognition from a Domain Adaptation Perspective M. Abdullah Jamal Matthew Brown Ming-Hsuan Yang Liqiang Wang Boqing Gong

Long-tailed Problem Emerging challenge as the datasets grow in scale Prevalent in fine-grained recognition, detection, etc. Datasets: iNaturalist, LVIS, ImageNet, COCO, etc. Vi Visual Genome

Shortcomings of Current Approaches Accuracy Accu cy on n Hea ead Classes es Ac Accuracy on on Head Classes Ac Accuracy on on Head Classes Accuracy on Ac on Tail Classes Accuracy on Ac on Tail Classes Accu Accuracy cy on n Tail Classes es

New Perspective - Domain Adaptation Slide source

Existing Works Assume target shift 𝜭 s (x|Common Slider) = 𝜭 t (x|Common Slider) 𝜭 s (x|King Eider) = 𝜭 t (x|King Eider)

But 𝜭 s (x|Common Slider) = 𝜭 t (x|Common Slider) 𝜭 s (x|King Eider) ≠ 𝜭 t (x|King Eider)

A Bird’s Eye View Example weights Training Stage w ƒ(x; 𝜄 ) ℒ Training Loss Inference Stage ƒ(x; 𝜄 ) Expects to perform well on all classes

Two-Component Approach [CVPR’19] Class-Balanced Loss Based on Effective Number of Samples [ICML’18] Learning to reweight examples for robust deep learning (1 - 𝞬 ) / ( 1- 𝞬 n ) Meta-learning framework

Two-Component Approach L2RW Ours ✓ Pre-training X ✓ Clip negative 𝝑 X ✓ Normalization X Free Space of 𝝑 reduced larger [CVPR’19] Class-Balanced Loss Based on Effective Number of Samples [ICML’18] Learning to reweight examples for robust deep learning (1 - 𝞬 ) / ( 1- 𝞬 n ) Meta-learning framework

Experiments Six datasets ● CIFAR-LT-10 ● CIFAR-LT-100 ● iNaturalist 2017 & 2018 ● ImageNet-LT ● Places-LT

CIFAR-LT-10 - Results

CIFAR-LT-10 - Results

CIFAR-LT-10 - Results

CIFAR-LT-10 - Results

What are the learned 𝝑

Long-tailed visual Domain Adaptation recognition A powerhouse of ideas & techniques - A new perspective from Domain Adaptation - A two-component approach - Domain-invariant representations - Maximum Mean Discrepancy - SOTA results on six datasets - Curriculum Domain Adaptation - Adversarial adaptation - Self-supervised adaptation