/Problem-Summary-Vehicle-Pose-Transform-by-GAN-for-Re-Id-KeyPoint-generator

Discuss and record the summary about "Vehicle Pose Transform by GAN for Re-Id KeyPoint generator" problem.

Problem Summary: Vehicle Pose Transform by GAN for Re-Id KeyPoint generator

This is a discuss and record for the summary about "Vehicle Pose Transform by GAN for Re-Id KeyPoint generator" problem. Now is working on the Sol.2., create another model to satisfy the GAN framework.

Vehicle Pose Transform by GAN for Re-Id

The structure of system framework is below:

圖片1

With training

  1. will input image Input Image x and a sequence with vehicle key-point Target pose.
  2. Generator generate fake image with target pose in specific key-point, target pose is come from real image.
  3. Compute Loss between fake image and real image, real image includes the target pose.

With testing

  1. will input image Input Image x without vehicle key-point Target pose.
  2. Generator generate fake image in target pose with specific key-point.

Problem Summary

Problem

In training, not all vehicle have all type of pose, so Generator will lack some specific key-point. we can not generate doesn't exist pose vehicle. The structure of training framework will become:

圖片

without key-point, generator will fail to generate fake image.

Solution

  1. Take the average of the KeyPoints for each pose of all types, and use the average instead if a vehicle of a certain type does not have KeyPoint on a specific pose.
  2. Create another model for generate vehicle KeyPoint.
  3. If vehicle doesn't have KeyPoint on a specific pose, then use CycleGAN-like algo. to generate fake image, Else using original GAN.