Loss Functions

# Dynamic SmoothL1 Loss

Introduced by Zhang et al. in Dynamic R-CNN: Towards High Quality Object Detection via Dynamic Training

Dynamic SmoothL1 Loss (DSL) is a loss function in object detection where we change the shape of loss function to gradually focus on high quality samples:

$$\text{DSL}\left(x, \beta_{now}\right) = 0.5|{x}|^{2}/\beta_{now}, \text{ if } |x| < \beta_{now}\text{,}$$ $$\text{DSL}\left(x, \beta_{now}\right) = |{x}| - 0.5\beta_{now}\text{, otherwise}$$

DSL will change the value of $\beta_{now}$ according to the statistics of regression errors which can reflect the localization accuracy. It was introduced as part of the Dynamic R-CNN model.

#### Papers

Paper Code Results Date Stars