Learning-Deep-Learning

Bag of Freebies for Training Object Detection Neural Networks

Mar 2019

tl;dr: Commonly used tricks to boost object detectors (both one and two-stage pipelines). Cosine learning rate, class label smoothing and mixup is very useful. Half-half mixup gives the largest performance boost.

Overall impression

Cosine learning rate, class label smoothing and mixup is very useful. It is surprising that mixup technic is useful in object detection setting. (Need more investigation into this topic)

Key ideas

Technical details

Notes