TL;DR
- task : uncertainty in FCOS model
- problem : I want to get uncertainty for 4 directions: left, right, top, bottom.
- idea : Assume 4 Gaussians
- architecture : FCOS + uncertainty branch + classification head with output from uncertainty branch in fusion
- objective : uncertainty loss : Put the IoU term as an exponential term in the KL divergence for the four Gaussians. cls loss : focal loss reflecting uncertainty
- baseline : centerness-branch, IoU branch, QFL, VFL
- data : COCO2017
- result : FCOS performance increased by 1.8 points
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
This is a good read for later
Architecture
Loss
Uncertainty Loss
Uncertainty Focal Loss
Result
