object-detection-metrics 0.3.post1

Object Detection Metrics

This project was forked from rafaelpadilla/Object-Detection-Metrics.

Getting started

Installing object_detection_metrics

$ pip install object_detection_metrics

Reading Josn file

import podm
bounding_boxes = podm.load_data('tests/sample_2/groundtruths.json')

Reading COCO file

import podm
bounding_boxes = podm.load_data_coco('tests/sample_2/groundtruths_coco.json')

PASCAL VOC Metrics

import podm
gt_BoundingBoxes = podm.load_data('tests/sample_2/groundtruths.json')
pd_BoundingBoxes = podm.load_data('tests/sample_2/detections.json')
results = podm.get_pascal_voc_metrics(gt_BoundingBoxes, pd_BoundingBoxes, .5)

ap, precision, recall, tp, fp, etc

for cls, metric in actuals.items():
    label = m.label
    print('ap', metric.ap)
    print('precision', metric.precision)
    print('interpolated_recall', metric.interpolated_recall)
    print('interpolated_precision', metric.interpolated_precision)
    print('tp', metric.tp)
    print('fp', metric.fp)
    print('num_groundtruth', metric.num_groundtruth)
    print('num_detection', metric.num_detection)

mAP

from podm import MetricPerClass
mAP = MetricPerClass.mAP(results)

IoU

box1 = Box(0., 0., 10., 10.)
box2 = Box(1., 1., 11., 11.)
Box.intersection_over_union(box1, box2)

Implemented metrics

Tutorial

• Intersection Over Union (IOU)
• TP and FP
  • True Positive (TP): IOU ≥ IOU threshold (default: 0.5)
  • False Positive (FP): IOU < IOU threshold (default: 0.5)
• Precision and Recall
• Average Precision
  • 11-point AP
  • all-point AP