mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Add evaluation to tk::dnn namespace, style fix also

Browse Source 
Signed-off-by: Micaela Verucchi <micaelaverucchi@gmail.com>
This commit is contained in:
Micaela Verucchi
2020-04-06 18:12:38 +02:00
parent c0a978a480
commit c8308963df
3 changed files with 81 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/evaluation.cpp
+67 -103
View File
@@ -1,12 +1,13 @@
#include "evaluation.h"
#include <fstream>
namespace tk { namespace dnn {
void BoundingBox::clear()
{
unique_truth_index = -1;
truth_flag = 0;
max_IoU = 0;
uniqueTruthIndex = -1;
truthFlag = 0;
maxIoU = 0;
}
bool boxComparison (const BoundingBox& a,const BoundingBox& b)
@@ -18,15 +19,15 @@ std::ostream& operator<<(std::ostream& os, const BoundingBox& bb)
{
os <<"w: "<< bb.w << ", h: "<< bb.h << ", x: "<< bb.x << ", y: "<< bb.y <<
", cat: "<< bb.cl << ", conf: "<< bb.prob<< ", truth: "<<
bb.truth_flag<< ", assignedGT: "<< bb.unique_truth_index<<
", maxIoU: "<< bb.max_IoU<<"\n";
bb.truthFlag<< ", assignedGT: "<< bb.uniqueTruthIndex<<
", maxIoU: "<< bb.maxIoU<<"\n";
return os;
}
void Frame::print() const
{
std::cout<<"labels filename: "<<l_filename<<std::endl;
std::cout<<"image filename: "<<i_filename<<std::endl;
std::cout<<"labels filename: "<<lFilename<<std::endl;
std::cout<<"image filename: "<<iFilename<<std::endl;
std::cout<<"GT: "<<std::endl;
for(auto g: gt) std::cout<<g;
std::cout<<"DET: "<<std::endl;
@@ -38,7 +39,7 @@ void PR::print()
std::cout<<"precision: "<<precision<<" recall: "<<recall<<" tp: "<<tp<<" fp:"<<fp<<" fn:"<<fn<<std::endl;
}
float overlap(float x1, float w1, float x2, float w2)
float boxOverlap(float x1, float w1, float x2, float w2)
{
float l1 = x1 - w1/2;
float l2 = x2 - w2/2;
@@ -51,8 +52,8 @@ float overlap(float x1, float w1, float x2, float w2)
float boxIntersection(const BoundingBox &a, const BoundingBox &b)
{
float w = overlap(a.x, a.w, b.x, b.w);
float h = overlap(a.y, a.h, b.y, b.h);
float w = boxOverlap(a.x, a.w, b.x, b.w);
float h = boxOverlap(a.y, a.h, b.y, b.h);
if(w < 0 || h < 0)
return 0;
float area = w*h;
@@ -69,15 +70,13 @@ float boxUnion(const BoundingBox &a, const BoundingBox &b)
float boxIoU(const BoundingBox &a, const BoundingBox &b)
{
float I = boxIntersection(a, b);
// std::cout<<"I: "<<I<<std::endl;
float U = boxUnion(a, b);
// std::cout<<"U: "<<U<<std::endl;
if (I == 0 || U == 0)
return 0;
return I / U;
}
void readParams(char* config_filename, int& classes, int& map_points,
void readmAPParams(char* config_filename, int& classes, int& map_points,
int& map_levels, float& map_step, float& IoU_thresh,
float& conf_thresh, bool& verbose)
{
@@ -106,8 +105,7 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
std::vector<int> dets_classes_count(classes,0);
//count groundtruth and detections in total and for each class
for(auto i:images)
{
for(auto i:images){
for(auto gt:i.gt)
truth_classes_count[gt.cl]++;
for(auto det:i.det)
@@ -116,8 +114,7 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
groundtruths_count += i.gt.size();
}
if(verbose)
{
if(verbose){
std::cout<<"gt_count: "<<groundtruths_count<<std::endl;
std::cout<<"det_count: "<<detections_count<<std::endl;
}
@@ -129,30 +126,24 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
// for each detection comput IoU with groundtruth and match detetcion and
// groundtruth with IoU greater than IoU_thresh
for(auto &img:images)
{
for(size_t i=0; i<img.det.size(); i++)
{
if(img.det[i].prob > conf_thresh)
{
for(auto &img:images){
for(size_t i=0; i<img.det.size(); i++){
if(img.det[i].prob > conf_thresh){
float maxIoU = 0;
int truth_index = -1;
for(size_t j=0; j<img.gt.size(); j++)
{
for(size_t j=0; j<img.gt.size(); j++){
float currentIoU = boxIoU(img.det[i], img.gt[j]);
if(currentIoU > maxIoU && img.det[i].cl == img.gt[j].cl)
{
if(currentIoU > maxIoU && img.det[i].cl == img.gt[j].cl){
maxIoU = currentIoU;
truth_index = j;
}
}
// std::cout<<"det i:"<<i<<" maxIoU:"<<maxIoU<<" tIndex:"<<truth_index<<std::endl;
if(truth_index > -1 && maxIoU > IoU_thresh)
{
if(truth_index > -1 && maxIoU > IoU_thresh){
// std::cout<<"(INSIDE) IoU thresh:"<<IoU_thresh<<" maxIoU:"<<maxIoU<<" maxIoU > IoU_thresh:"<<(maxIoU > IoU_thresh)<<std::endl;
img.det[i].unique_truth_index = truth_index + gt_checked;
img.det[i].truth_flag = 1;
img.det[i].max_IoU = maxIoU;
img.det[i].uniqueTruthIndex = truth_index + gt_checked;
img.det[i].truthFlag = 1;
img.det[i].maxIoU = maxIoU;
}
}
@@ -161,8 +152,7 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
gt_checked += img.gt.size();
}
if(verbose)
{
if(verbose){
for(auto img:images)
img.print();
std::cout<<"\n\n\n\n";
@@ -178,30 +168,24 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
//compute precision-recall curve
std::vector<std::vector<PR>> pr( classes, std::vector<PR>(detections_count));
for(int rank = 0; rank< detections_count; ++rank)
{
if (rank > 0)
{
for (int class_id = 0; class_id < classes; ++class_id)
{
for(int rank = 0; rank< detections_count; ++rank){
if (rank > 0) {
for (int class_id = 0; class_id < classes; ++class_id) {
pr[class_id][rank].tp = pr[class_id][rank - 1].tp;
pr[class_id][rank].fp = pr[class_id][rank - 1].fp;
}
}
//if it was detected and never detected before
if (all_dets[rank].truth_flag == 1 && truth_flags[all_dets[rank].unique_truth_index] == 0)
{
truth_flags[all_dets[rank].unique_truth_index] = 1;
if (all_dets[rank].truthFlag == 1 && truth_flags[all_dets[rank].uniqueTruthIndex] == 0) {
truth_flags[all_dets[rank].uniqueTruthIndex] = 1;
pr[all_dets[rank].cl][rank].tp++; // true-positive
}
else
{
else {
pr[all_dets[rank].cl][rank].fp++; // false-positive
}
for (int i = 0; i < classes; ++i)
{
for (int i = 0; i < classes; ++i){
const int tp = pr[i][rank].tp;
const int fp = pr[i][rank].fp;
const int fn = truth_classes_count[i] - tp; // false-negative = objects - true-positive
@@ -217,17 +201,15 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
else
pr[i][rank].recall = 0;
if (rank == (detections_count - 1) && dets_classes_count[i] != (tp + fp))
{ // check for last rank
if (rank == (detections_count - 1) && dets_classes_count[i] != (tp + fp)) {
// check for last rank
printf(" class_id: %d - detections = %d, tp+fp = %d, tp = %d, fp = %d \n", i, dets_classes_count[i], tp+fp, tp, fp);
}
}
}
if(verbose)
{
for(int i=0; i < pr.size(); i++)
{
if(verbose){
for(int i=0; i < pr.size(); i++) {
std::cout<<"---------Class "<<i<<std::endl;
for(auto r:pr[i])
r.print();
@@ -240,16 +222,13 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
double last_recall, last_precision, delta_recall;
double cur_recall, cur_precision;
double avg_precision = 0;
for (int i = 0; i < classes; ++i)
{
for (int i = 0; i < classes; ++i) {
avg_precision = 0;
if (map_points == 0) //mAP calculation: ImageNet, PascalVOC 2010-2012
{
if (map_points == 0){ //mAP calculation: ImageNet, PascalVOC 2010-2012
last_recall = pr[i][detections_count - 1].recall;
last_precision = pr[i][detections_count - 1].precision;
for (int rank = detections_count - 2; rank >= 0; --rank)
{
for (int rank = detections_count - 2; rank >= 0; --rank){
delta_recall = last_recall - pr[i][rank].recall;
last_recall = pr[i][rank].recall;
@@ -259,8 +238,7 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
avg_precision += delta_recall * last_precision;
}
}
else //MSCOCO - 101 Recall-points, PascalVOC - 11 Recall-points
{
else {//MSCOCO - 101 Recall-points, PascalVOC - 11 Recall-points
for (int point = 0; point < map_points; ++point) {
cur_recall = point * 1.0 / ( map_points - 1 );
cur_precision = 0;
@@ -287,16 +265,14 @@ double computeMap(std::vector<Frame> &images,const int classes,const float IoU_t
double computeMapNIoULevels(std::vector<Frame> &images,const int classes,const float i_IoU_thresh, const float conf_thresh, const int map_points, const float map_step, const int map_levels, const bool verbose, const bool write_on_file, std::string net)
{
std::ofstream out_file;
if(write_on_file)
{
if(write_on_file){
out_file.open("map.csv", std::ios_base::app);
out_file<<net<<";";
}
double AP = 0, cur_AP = 0;
float IoU_thresh = i_IoU_thresh;
for(int i=0; i<map_levels; ++i)
{
for(int i=0; i<map_levels; ++i){
for(auto& img:images)
for(auto & d:img.det)
d.clear();
@@ -308,10 +284,9 @@ double computeMapNIoULevels(std::vector<Frame> &images,const int classes,const f
}
AP/=map_levels;
if(write_on_file)
{
out_file<<AP<<"\n";
out_file.close();
if(write_on_file){
out_file<<AP<<"\n";
out_file.close();
}
return AP;
}
@@ -320,9 +295,8 @@ void computeTPFPFN(std::vector<Frame> &images,const int classes,const float IoU_
{
std::ofstream out_file;
if(write_on_file)
{
out_file.open("pr.csv", std::ios_base::app);
if(write_on_file){
out_file.open("pr.csv", std::ios_base::app);
out_file<<net<<";";
}
@@ -330,50 +304,42 @@ void computeTPFPFN(std::vector<Frame> &images,const int classes,const float IoU_
std::vector<int> dets_classes_count(classes,0);
std::vector<PR> pr(classes);
for(auto &img:images)
{
for(auto &img:images){
for(auto& tc: truth_classes_count)
tc = 0;
for(auto& dc: dets_classes_count)
dc = 0;
std::vector<bool> det_assigned(img.det.size(), false);
for(size_t j=0; j<img.gt.size(); j++)
{
for(size_t j=0; j<img.gt.size(); j++){
truth_classes_count[img.gt[j].cl]++;
float maxIoU = 0;
int det_index = -1;
for(size_t i=0; i<img.det.size(); i++)
{
if(img.det[i].prob > conf_thresh)
{
for(size_t i=0; i<img.det.size(); i++){
if(img.det[i].prob > conf_thresh){
float currentIoU = boxIoU(img.det[i], img.gt[j]);
if(currentIoU > maxIoU && img.det[i].cl == img.gt[j].cl && !det_assigned[i])
{
if(currentIoU > maxIoU && img.det[i].cl == img.gt[j].cl && !det_assigned[i]){
maxIoU = currentIoU;
det_index = i;
}
}
}
if(det_index > -1 && maxIoU > IoU_thresh && !det_assigned[det_index])
{
img.det[det_index].unique_truth_index = j;
img.det[det_index].truth_flag = 1;
img.det[det_index].max_IoU = maxIoU;
if(det_index > -1 && maxIoU > IoU_thresh && !det_assigned[det_index]){
img.det[det_index].uniqueTruthIndex = j;
img.det[det_index].truthFlag = 1;
img.det[det_index].maxIoU = maxIoU;
det_assigned[det_index] = true;
dets_classes_count[img.det[det_index].cl]++;
}
}
for(size_t i=0; i<img.det.size(); i++)
{
if(img.det[i].truth_flag)
for(size_t i=0; i<img.det.size(); i++){
if(img.det[i].truthFlag)
pr[img.det[i].cl].tp ++;
else
pr[img.det[i].cl].fp ++;
}
for(size_t i=0; i<classes; i++)
{
for(size_t i=0; i<classes; i++){
pr[i].fn += truth_classes_count[i] - dets_classes_count[i];
}
}
@@ -382,8 +348,7 @@ void computeTPFPFN(std::vector<Frame> &images,const int classes,const float IoU_
int TP = 0, FP = 0, FN = 0;
for(size_t i=0; i<classes; i++)
{
for(size_t i=0; i<classes; i++){
pr[i].precision = (pr[i].tp + pr[i].fp) > 0 ? (double)pr[i].tp / (double)(pr[i].tp +pr[i].fp) : 0;
pr[i].recall = (pr[i].tp + pr[i].fn) > 0 ? (double)pr[i].tp / (double)(pr[i].tp +pr[i].fn) : 0;
if(verbose)
@@ -392,22 +357,21 @@ void computeTPFPFN(std::vector<Frame> &images,const int classes,const float IoU_
avg_precision += pr[i].precision;
avg_recall += pr[i].recall;
TP += pr[i].tp;
FP += pr[i].fp;
FN += pr[i].fn;
TP += pr[i].tp;
FP += pr[i].fp;
FN += pr[i].fn;
}
avg_precision /= classes;
avg_recall /= classes;
f1_score = avg_precision + avg_recall > 0 ? 2 * ( avg_precision * avg_recall ) / ( avg_precision + avg_recall ) : 0;
if(write_on_file)
{
out_file<<TP<<";"<<FP<<";"<<FN<<";"<<avg_precision<<";"<<avg_recall<<";"<<f1_score<<"\n";
out_file.close();
if(write_on_file){
out_file<<TP<<";"<<FP<<";"<<FN<<";"<<avg_precision<<";"<<avg_recall<<";"<<f1_score<<"\n";
out_file.close();
}
std::cout<<"avg precision: "<<avg_precision<<"\tavg recall: "<<avg_recall<<"\tavg f1 score:"<<f1_score<<std::endl;
}
}}