paper: Fast R-CNN

SPPnet solved the R-CNN’s problem that it extracts features for each of the 2k~ region proposal and costs a lot of time. SPPnet runs the convolutional layers only once on the entire image (regardless of the number of windows), and then extract features by SPP-net on the feature maps. But it is not an end-to-end model so that extracted features need to be written to the disk, and uses two stages for classification and bbox regression. Fast R-CNN can do back-propagation end-to-end.

Architecture


I think the main difference of Fast R-CNN is that it’s RoI layer is simply the special-case of the spatial pyramid pooling layer used in SPPnets in which there is only one pyramid level to get fixed feature map. Then using a single-stage training algorithm that jointly learns to classify object proposals and refine their spatial locations. The loss function is as follows:

$L\left(p, u, t^{u}, v\right)=L_{\mathrm{cls}}(p, u)+\lambda[u \geq 1] L_{\mathrm{loc}}\left(t^{u}, v\right)$
classification loss use cross-entropy and regression loss use smooth L1 loss. However, it still use selective search to achieve region proposals.

Others

  • Fast R-CNN uses a faster sampling method for training compare to R-CNN and SPPnets, more details are in paper.
  • use truncated SVD to compress the fc layers.