[78] Localization Uncertainty Estimation for Anchor-Free Object Detection

[long8v]

paper

TL;DR

• task : uncertainty in FCOS model
• problem : left, right, top, bottom 4 방향에 대한 uncertainty를 얻고 싶다.
• idea : 4개의 가우시안을 가정하자
• architecture : FCOS + uncertainty branch + classification head에 uncertainty branch의 output을 fusion
• objective : uncertainty loss : 4개의 가우시안에 대한 KL divergence에 IoU term을 지수항으로 넣어줌. cls loss : uncertainty를 반영한 focal loss
• baseline : centerness-branch, IoU branch, QFL, VFL
• data : COCO2017
• result : FCOS 성능 1.8 point 올림

Details

Related work

• generalized focal loss https://arxiv.org/pdf/2006.04388.pdf

• FCOS: Fully Convolutional One-Stage Object Detection https://arxiv.org/pdf/1904.01355.pdf 이건 나중에 한번 읽어보면 좋을듯

Architecture

Loss

Uncertainty Loss

Uncertainty Focal Loss

Result