If we're to be specific... From thread:

Following software runs now on the N900:
- ROS
- OpenCV
- gstreamer (640x480 H.264)
- fcam (used intitially for autofocus etc)
- VNC / SSH (for remote access)

I wrote an ROS node for the N900 accelerometer and played a little bit with OpenCV. The face detection speed on the N900 was a bit disappointing though (0.25 fps) and I decided to stream the video to my PC (using gstreamer for which exists an ROS integration). Streaming causes only about 20% CPU utilization (640x480 resolution / H.264 / 1500 KB/s) on the N900.
Less complex tasks like color based tracking should be able to run on the robot itself with acceptable speed.

Computational tasks (distance analysis from cam feedback) most likely were run from a desktop PC, even at 1150mhz N900 alone most likely couldn't handle it.

Just noticed: That or what MINKIN2 pointed out, possibly both