Please post your tf tree.

link text this is the tf tree generated after executing rosrun tf view_frames

If you are gathering point cloud in one of _optical frames, z coordinate should point forward from the camera. If not, the driver you use may be written disregarding REP-103.

so you mean that z-coordinate would itself give me depth? what would be the units then?

No I don't. To be able to "mean" anything, you need to supply more details (like sensor model, driver you use, topic you listen and message samples etc.). Drivers should follow REP-103 for units but they do not have to.

Are you wanting radial distance, like if you had a ruler from the center of your camera to the object - distance = sqrt(x^2 + y^2 + z^2)?

Yes it is what I need!!!! @lucasw

I don't think there is an existing node to do that for you- is implementing that equation in your own node unclear or is just having the equation a sufficient answer to this question?

To implement it in my code I need to know xyz coordinates of the object of whose depth I want. But how do I get those coordinates??

If you are asking how to do object recognition check out the pcl tutorials, http://pointclouds.org/documentation/tutorials/ , or if you are asking how to iterate through a pointcloud to get xyz values look at http://wiki.ros.org/pcl_ros#Subscribing_to_point_clouds