Decoupling Representation and Classifier for Long-Tailed Recognition, Kang, Xie, Rohrbach, Yan, Gordo, Feng, Kalantidis; 2019 - Summary
author: joshpapermaster
score: 8 / 10

What is the core idea?

Visual recognition models can struggle when there is a long-tailed distribution. Most of the current solutions to the long-tailed problem jointly learn feature representations and train classifiers. This paper decouples their solutions into two processes: representation learning and classification. This is beneficial because they are able to determine which techniques are specifically effective.

Representation learning

Classification

The paper determined the following findings from using their decoupled process:

How is it realized technically?

Experiments covered the following different decoupled techniques to determine which were most effective

Representation Learning:

decoupling

Probability of sampling an image from class j where C is the number of classes

Classification:

How well does the paper perform?

Experiments trained on three datasets

The models were then tested on the respective balanced dataset over all classes

The many, medium, few, and all titles below refer to the number of training images the model had for that the class. e.g. the results under “Few” show how well the model performed when the training set only contained less than 20 images

Sampling method critically matters for joint methods, but overall decoupled methods worked better. decoupling

Looking at different areas of restarting learning decoupling

Instance-balanced sampling provides generalizable representations decoupling

Results continue to show decoupled methods providing the best results decoupling decoupling

What interesting variants are explored?

TL;DR