On Label Granularity and Object Localization

TL;DR

This paper investigates how the level of label detail affects weakly supervised object localization, introducing a new dataset and showing that label granularity has a greater impact on performance than algorithm choice.

Contribution

It introduces iNatLoc500, a large-scale fine-grained dataset for WSOL, and demonstrates the significant influence of label granularity on localization performance and data efficiency.

Findings

Choosing the right label granularity greatly improves WSOL performance.

Label granularity impacts data efficiency more than algorithm selection.

A new fine-grained dataset, iNatLoc500, is introduced for WSOL research.

Abstract

Weakly supervised object localization (WSOL) aims to learn representations that encode object location using only image-level category labels. However, many objects can be labeled at different levels of granularity. Is it an animal, a bird, or a great horned owl? Which image-level labels should we use? In this paper we study the role of label granularity in WSOL. To facilitate this investigation we introduce iNatLoc500, a new large-scale fine-grained benchmark dataset for WSOL. Surprisingly, we find that choosing the right training label granularity provides a much larger performance boost than choosing the best WSOL algorithm. We also show that changing the label granularity can significantly improve data efficiency.