CaltechAUTHORS
  A Caltech Library Service

Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed Detection

Chang, Nadine and Yu, Zhiding and Wang, Yu-Xiong and Anandkumar, Animashree and Fidler, Sanja and Alvarez, Jose M. (2021) Image-Level or Object-Level? A Tale of Two Resampling Strategies for Long-Tailed Detection. Proceedings of Machine Learning Research, 139 . pp. 1463-1472. ISSN 2640-3498. https://resolver.caltech.edu/CaltechAUTHORS:20210510-134322482

[img] PDF - Published Version
See Usage Policy.

3MB
[img] PDF - Submitted Version
See Usage Policy.

3MB

Use this Persistent URL to link to this item: https://resolver.caltech.edu/CaltechAUTHORS:20210510-134322482

Abstract

Training on datasets with long-tailed distributions has been challenging for major recognition tasks such as classification and detection. To deal with this challenge, image resampling is typically introduced as a simple but effective approach. However, we observe that long-tailed detection differs from classification since multiple classes may be present in one image. As a result, image resampling alone is not enough to yield a sufficiently balanced distribution at the object-level. We address object-level resampling by introducing an object-centric sampling strategy based on a dynamic, episodic memory bank. Our proposed strategy has two benefits: 1) convenient object-level resampling without significant extra computation, and 2) implicit feature-level augmentation from model updates. We show that image-level and object-level resamplings are both important, and thus unify them with a joint resampling strategy. Our method achieves state-of-the-art performance on the rare categories of LVIS, with 1.89% and 3.13% relative improvements over Forest R-CNN on detection and instance segmentation.


Item Type:Article
Related URLs:
URLURL TypeDescription
https://proceedings.mlr.press/v139/chang21c.htmlPublisherArticle
https://arxiv.org/abs/2104.05702arXivDiscussion Paper
Additional Information:© 2021 by the author(s). We would like to sincerely thank Achal Dave, Kenneth Marino, Senthil Purushwalkam and other NVIDIA colleagues for the discussion and constructive suggestions.
Record Number:CaltechAUTHORS:20210510-134322482
Persistent URL:https://resolver.caltech.edu/CaltechAUTHORS:20210510-134322482
Usage Policy:No commercial reproduction, distribution, display or performance rights in this work are provided.
ID Code:109038
Collection:CaltechAUTHORS
Deposited By: Tony Diaz
Deposited On:10 May 2021 20:56
Last Modified:04 Oct 2021 18:23

Repository Staff Only: item control page