Menu

Evaluation

Agostino Gibaldi
Attachments
accuracy.png (49085 bytes)
normalization.png (73588 bytes)
robot_heads.png (75024 bytes)
vergence_range.png (107134 bytes)
- Overview and credits
- How to install
- Demo video
- Code explanation
- Binocular energy model
- Vergence control
- Control Evaluation

Control Evaluation

In order to show the portability of the control, it was tested on three robotic heads with different kinematic characteristics: iCub, Searise and Koala.

The effectiveness has been evaluated using three different and complementary principles:
1. accuracy and precision
2. operative range
3. robustness to real-world conditions

robot_heads

Accuracy and Precision

A vergence movement towards a frontoparallel surface was repeated 250 times. The target plane was at a fixed vergence distance of 8 degrees, and the binocular fixation point started at a random vergence distance between 4 and 12 degrees. The accuracy has been measured as the residual disparity at the end of the vergence movement, and displayed as mean with respect to the starting vergence position.

accuracy

Operative Range

While exploring the environment, the robot agent needs to move the fixation point from close to far targets and viceversa in the peripersonal space and farther.
Starting from a reference fixation distance of 8◦ of vergence, we moved a surface perpendicular to the gaze direction from far to close distances several times keeping the fixation point steady. The range of effectiveness is defined by themaximum (farthest) and minimum (closest) distance from the fixation point where the control is able to produce the correct vergence movement.

The figure shows the range of effectiveness measured on the iCub stereo head.

vergence_range

Robustness to Real-World Conditions

To quantitatively assess the performance of our approach under variations of external conditions, we can derive an approximation of the disparity-vergence curves associated to the control signal. To evaluate the robustness of the control to real-worls lighting conditions, the curve was then computed considering two conditions:
1. using different luminous powers (4400 and 6600 lm)
2. changing the interocular contrast between the left and the right images.

normalization


Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.