What must be the scale of point cloud data to be published in costmap ?

edit

ROS uses SI units for all its messages so pointclouds should be described in meters at all times.

To answer your question about converting from image space (pixels) to world space (m) I think you'll want to know about Camera calibration. This is the process of measuring a camera lens so that you know exactly what angle a pixel represents.

Converting calibrated images to 3D information is a very large and complex area. I recommend you read up on Stereopsis and structure from motion to get a general understanding.