In R-CNN paper, they give the definition of the target values for bounding box regression

Given that $(P, G)$ is a (prediction box, ground-truth box) pair of the form $(x, y, w, h)$ where $x, y$ is the center coordinate of the box, $w, h$ are width and height respectively.

$t_x = (G_x - P_x) / P_w \hspace{2.0cm} t_y = (G_y - P_y) / P_h$

$t_w = \log(G_w / P_w) \hspace{2.0cm} t_h = \log(G_h / P_h)$

And the goal is to find $\textbf{w}_*$, where $*$ can be $x, y, w$ or $h$, so that

$\textbf{w}_* = \arg \min_{\hat{\textbf{w}}_*} \sum_i (t^i_* - \hat{\textbf{w}}_*^T \phi(P^i))^2 + \lambda \|\hat{\textbf{w}}_*\|^2$ where $\phi(P^i)$ is the feature map given by the last pooling layer of the feature extractor after passing predicted bounding box $P^i$

I don't understand why they come up with this approach of bounding box regression ? Can anyone tell me about this ?

P.S: since this regression approach is used not only in R-CNN but also in later models, I really want to get a clear understanding of this


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.