Evaluation

Sub-module containing classes and methods for evaluation the classification results. More...

Collaboration diagram for Evaluation:

Classes
class	DirectGraphicalModels::CCMat
	Confusion matrix class. More...

Functions
float	DirectGraphicalModels::getAveragePrecision (const vec_byte_t &predictions, const vec_float_t &potentials, const vec_byte_t &gt, byte state)
	Returns Average Precision for the selected state (class) state. More...

Detailed Description

Sub-module containing classes and methods for evaluation the classification results.

Function Documentation

getAveragePrecision()

float DirectGraphicalModels::getAveragePrecision	(	const vec_byte_t &	predictions,
		const vec_float_t &	potentials,
		const vec_byte_t &	gt,
		byte	state
	)

Returns Average Precision for the selected state (class) state.

This function analyses 3 values of every image pixel, namely the predicted state, the groundtruth state and the potential of the pixel to be the class state, passed to the function via the arguments: predictions, gt and potentials, respectively. First, the pixels are sorted by the potentials in descending order. After that, the following algorithm is applied:

for (int i = 1; i <= numPixels; i++) 
  if (Pixel[i].groundtruth_state == state) {                
    numRelevants++;
    if (Pixel[i].predicted_state == state) {            
        numCoincidences++;
        AP = AP + numCoincidences / i;
    }
  }
AP = AP / nRelevants;

where AP stays for Average Precision.

Parameters

predictions	The most probable configuration, returned by the CDecode::decode() function
potentials	The potential values for each node of the graph, returned by the CInfer::getPotentials(state) function
gt	The groundtruth values for each node of the graph. May be converted from a groundtruth image as follows: vec_byte_t gt(gtImg.data, gtImg.data + gtImg.cols * gtImg.rows);
state	The state (class) for which the Average Precision is calculated

Returns

The Average Precision value

Definition at line 6 of file AveragePrecision.cpp.