Learning-AI

Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection

December 2020

tl;dr: Introduces a bridge between QFL and DFL in generalized focal loss.

Overall impression

The improvement of GFocalV2 over GFocal is simple. GFocal predicts QFL (quality focal loss) and DFL (distribution focal loss) separately, while GFocalV2 introduced a bridge named distribution guided quality predictor (DGQP) to guide the prediction of localization quality estimation (LQE).

The classification and localization still has a joint representation, but the training and prediction method has a decomposed design.

The acronym in this manuscript is getting a bit crazy. Even in the original GFocal paper, the annotation is a bit unnecessarily obscure.

Key ideas

Technical details

Notes