D-DL4CV-Lec15c-R-CNN

R-CNN

Introduction

R-CNN is a kind of multiple objects detection method that leverage the concept of “region proposals”

Steps

Step 1: Region Proposals

Run region proposal method to compute ~2000 candidates

Step 2: Make Predictions

Resize each region to size of CNN’s input, then pass them through CNN to predict class scores and bounded box transform

Step 3: Select Output

In this step we select a subset of region proposals to output based on the predict scores and our choice of choosing method

Step 4: Compare boxes

We use Intersection over Union (IoU) to compare our prediction with the ground-truth box which is drawn manually

Bounded Box Transform Prediction

Explanation

This part of the output predicts the “transform” to the box in order to correct the box size and position

A Common Practice

We predict the “Transform”

Region Proposal: $(p_x,p_y,p_h,p_w)$ Transform: $(t_x,t_y,t_h,t_w)$ Output Box: $(b_x,b_y,b_h,b_w)$

Translate relative to the box:

$$b_x=p_x+p_w{t}_x{b}_y=p_y+p_h{t}_y$$

Log-space scale transform:

$$b_w=p_w\exp (t_w)b_h=p_h\exp (t_h)$$

Overlapping Boxes

Problem

Object detectors often output many bounding boxes that detect the same object multiple time

Solution: Non-Max Suppression (NMS)

D-DL4CV-Lec15cb-NMS

Evaluating Object Detectors

Problem

Object detection is very different from image classification, thus we need to come up with a new way evaluating how good our object detector is

Solution: Mean Average Precision (mAP)

D-DL4CV-Lec15cc-mAP

Problem of R-CNN

It is very slow since we need to forward pass for each image