Large-Scale Long-Tailed Recognition in an Open World Ziwei Liu* Zhongqi Miao* Xiaohang Zhan Jiayun Wang Boqing Gong Stella X. Yu The Chinese University of Hong Kong UC Berkeley / ICSI
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Large-Scale Long-Tailed Recognition in an Open World
Ziwei Liu* Zhongqi Miao* Xiaohang Zhan Jiayun Wang Boqing Gong Stella X. Yu
The Chinese University of Hong Kong UC Berkeley / ICSI
Train
Test
Cat
Fox
Panda
CatFoxPanda
CatFoxPanda
Train
Test
Cat
Fox
Panda
CatFoxPanda
CatFoxPanda
?(open class)
(many-shotclass)
(medium-shotclass)
(few-shotclass)
Faces [Zhang et al. 2017] Places [Wang et al. 2017]
Actions [Zhang et al. 2019]Species [Van Horn et al. 2019]
?Open Long-Tailed Recognition
Open World
Head Classes Tail Classes Open Classes
?Open Long-Tailed Recognition
Open World
Head Classes Tail Classes Open Classes
Knowledge Transfer
?Open Long-Tailed Recognition
Open World
Head Classes Tail Classes Open Classes
Sensitivity to NoveltyKnowledge Transfer
?Open Long-Tailed Recognition
Open World
Head Classes Tail Classes Open Classes
Avoid Forgetting
Sensitivity to NoveltyKnowledge Transfer
?Open Long-Tailed Recognition
Open World
Imbalanced Classification
Few-shot Learning
Head Classes Tail Classes Open Classes
Open Set Recognition
Imbalanced Classification(metric learning, re-sampling, re-weighting)
test train testtrain
Few-Shot Learning(meta learning, classifier dynamics)
Open Set Recognition(distribution rectification, out-of-distribution detection)
Open Long-Tailed Recognition(dynamic meta-embedding)
testtrain train test
Sensitivity to Novelty Avoid Forgetting
Knowledge Transfer Knowledge Transfer Sensitivity to Novelty
Avoid Forgetting
train test
Open Long-Tailed Recognition(dynamic meta-embedding)
Knowledge Transfer Sensitivity to Novelty
Avoid Forgetting
FLY
visual memory
top-down attention
enhanced embedding
FLY
familiarity
direct embedding
bottom-up attention
Head Classes Tail Classes
bottom-up attention
top-down attention
familiarity
visual memory
Open Classes
Head Classes Tail Classes
bottom-up attention
top-down attention
familiarity
visual memory
Open Classes
Avoid Forgetting Knowledge Transfer
Sensitivity to Novelty
Head Classes Tail Classes
bottom-up attention
Open Classes
top-down attention
familiarity
visual memory
originalfeature map
attentivefeature map
TenchHand
Fish
Head Classes Tail Classes
bottom-up attention
top-down attention
familiarity
visual memory
Open Classes
direct embedding
enhanced embedding
associative memory
feature selection
Head Classes Tail Classes
familiarity
visual memory
Open Classes
bottom-up attention
embedding
rescaled embedding
Tail Class ‘African Grey’
Tail Class ‘Buckeye’
top-down attention
Open Sample
ImageNet-LT Benchmark
Absolute Performance Gain: ~20%
Places-LT Benchmark
MS1M-LT Benchmark
Absolute Performance Gain: ~10%
Absolute Performance Gain: ~2%
Methods ImageNet-LT Places-LT MS1M-LTPlain Model 0.295 0.366 0.738Sample Re-weighting (Focal Loss) 0.371 0.453 -Metric Learning (Range Loss) 0.373 0.457 0.722Open Set Recognition (OpenMax) 0.368 0.458 -Few-shot Learning (FSLwF) 0.347 0.375 -
Dynamic Meta-Embedding 0.474 0.464 0.745
Overall F1 Score on ImageNet-LT, Places-LT and MS1M-LT Benchmarks
Methods ImageNet-LT Places-LT MS1M-LTPlain Model 0.295 0.366 0.738Sample Re-weighting (Focal Loss) 0.371 0.453 -Metric Learning (Range Loss) 0.373 0.457 0.722Open Set Recognition (OpenMax) 0.368 0.458 -Few-shot Learning (FSLwF) 0.347 0.375 -
Dynamic Meta-Embedding 0.474 0.464 0.745
Overall F1 Score on ImageNet-LT, Places-LT and MS1M-LT Benchmarks
Methods ImageNet-LT Places-LT MS1M-LTPlain Model 0.295 0.366 0.738Sample Re-weighting (Focal Loss) 0.371 0.453 -Metric Learning (Range Loss) 0.373 0.457 0.722Open Set Recognition (OpenMax) 0.368 0.458 -Few-shot Learning (FSLwF) 0.347 0.375 -
Dynamic Meta-Embedding 0.474 0.464 0.745
Overall F1 Score on ImageNet-LT, Places-LT and MS1M-LT Benchmarks
Methods ImageNet-LT Places-LT MS1M-LTPlain Model 0.295 0.366 0.738Sample Re-weighting (Focal Loss) 0.371 0.453 -Metric Learning (Range Loss) 0.373 0.457 0.722Open Set Recognition (OpenMax) 0.368 0.458 -Few-shot Learning (FSLwF) 0.347 0.375 -
Dynamic Meta-Embedding 0.474 0.464 0.745
Overall F1 Score on ImageNet-LT, Places-LT and MS1M-LT Benchmarks
Methods ImageNet-LT Places-LT MS1M-LTPlain Model 0.295 0.366 0.738Sample Re-weighting (Focal Loss) 0.371 0.453 -Metric Learning (Range Loss) 0.373 0.457 0.722Open Set Recognition (OpenMax) 0.368 0.458 -Few-shot Learning (FSLwF) 0.347 0.375 -
Dynamic Meta-Embedding 0.474 0.464 0.745
Overall F1 Score on ImageNet-LT, Places-LT and MS1M-LT Benchmarks
Few shot
New Task Open Long-Tailed Recognition (OLTR)
Train Test
New Approach Dynamic Meta-Embedding
New Benchmarks ImageNet-LT Places-LT MS1M-LT
Thanks! Code, models and benchmarks are available at
Project Page: https://liuziwei7.github.io/projects/LongTail.html
Poster #170
Related Documents
![arXiv:1803.06978v4 [cs.CV] 1 Jun 2019 · 2019. 6. 4. · Komodo dragon American alligator armadillo cat bear red fox pot walking stick proboscis monkey American alligator red fox](https://static.cupdf.com/doc/110x72/607abb5e21092068880ae98f/arxiv180306978v4-cscv-1-jun-2019-2019-6-4-komodo-dragon-american-alligator.jpg)