Faster-RCNN

Summary

0. History

Love between Kaiming&Ross v2, finally together, happy ever after

1. Objective

Reduce the region proposal computation time

2. Model adopted

Convolutional layers + RPN and RoI pooling layers + softmax classifier

3. Specialties of system

a. RPN takes an image of any size as input and outputs a set of rectangular object proposals, each with an objectness score. To generate region proposals, we use a sliding window over the output feature map(every region) of the last shared convolutional layer. Each window is mapped to a  lower-dimensional feature.

b. at each sliding window, we predict multiple region proposals, which we call anchors. We use the anchor boxes of multiple scales and sizes insteand of image with multiple scales

c. approch is translation invariant, both in term of the anchors and the functions that compute proposals relative to the anchors, which adapt to different scales and ratios of images, thanks to this proporty, the anchors can be applied for just once

d. anchors are multi-scale to save computation time

e. binary class labels are assigned to each anchor, an anchoris positive if it's IoU overlap is higher than 0.7 with any ground truth box or if max(IoU) with a ground truth box

f. using a minibatch of size 256 with the ratio between negative and positive examples equals to 1:1, to avoid the negative examples biasing the model

g. training process:

     - initialize the RPN with an ImageNet pre-trained model and fine tune it with region proposals task

     - same process for training detection network

     - use the detection net to initialize RPN training and fix the shared conv layers to only fine tune the layers unique to RPN

     - fix the conv layers and fine tune the layers unique of Fast-RCNN

     - more iterations contribute little improvement

h. during training, boundry anchor boxes are ignored and not taken account in the loss function, and anchor boxes are reduced thanks to NMS

i. top ranked RPN proposals are accurate

4. Personal reviews

Good intention, still can be modified in the section of RPN proposals

转载于:https://www.cnblogs.com/lucasdu/p/7884616.html