Probability Computation

Inputs
Probability Computation
Resulting Neural Network
Learning
Network Outputs
Numerical Results

Inputs

Inputs used to compute the probability are choosen among:

distance (splited or not) from initial condition (axis system, 3D database) to found detail.

lms distance(splited or not) from LMS axis system to found detail.
found detail energy.

examples: the vehicle is the one hypothesised.
counter examples: the vehicle is not the one hypothesised.

Examples

LMS distance

Detail energy

Counter Examples

LMS distance

Detail energy

Probability Computation

Resulting Neural Network

Learning

Network Outputs

The output is smooth, no overfitting has occured.
The shape is intuitively appropriate, our confidence is high if:
- The detail energy (grad energy) is high,
- The location is appropriate (low lms distance)
- The higher the energy, the more tolerant we are on the location.

The detail becomes less discriminant if its view angle increases:

Numerical Results

The grey cell is the correct match. For correct recognition, the probability in that cell should be higher than probabilities in other cells on the same line. Ideally, for good discriminating details, grey cell probability should be close to 1 while other should be close to 0.