Fast R-CNN

lastmod 2019-01-09created 2016-03-104 edits

Category

• Fast R-CNN:Paper

Fast R-CNN

The RoI pooling layer is a simplified version of the SPP layer. It takes only a single level scale. So with SPPnet (e.g., when using the VGG16 model), only the 3-layer softmax classifer can be fine-tuned, because the convoulutional features operate offline. Since Back-propagation doesn't work, only the first l3 layers are fixed at their initial values and only the last 3 layers are updated. So SPPnet can do the ROI-centric training developed in R-CNN. It computes over RoIs, and an RoI is usually about as large as the original image. So it is slow and uses too much memory. FRCN, however, is image-centric. mini-batches are sampled hierarchically: first images, then the RoIs of those images. RoIs from the same image share computation and memory.

The overall idea is similar to SPPnet, but it seems to make up for SPPnet's drawback of being unable to train the Conv layers by designing the Back Propogation algorithm in a better way.

Documentation

Fast R-CNN (Ross Girshick, Microsoft Research)

https://github.com/rbgirshick/fast-rcnn

Favorite site

• Microsoft Research: Fast R-CNN
• Fast R-CNN / Object detection with Caffe ¹
• [Recommended] Object Detection and Classification using R-CNNs

References