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

Development of object_detection_metrics happens on GitHub: https://github.com/yfpeng/object_detection_metrics

The latest object_detection_metrics releases are available over pypi.

Getting started

Installing object_detection_metrics

$ pip install object_detection_metrics

Reading COCO file

from podm import coco_decoder
with open('tests/sample/groundtruths_coco.json') as fp:
    gold_dataset = coco_decoder.load_true_object_detection_dataset(fp)

PASCAL VOC Metrics

from podm import coco_decoder
from podm.metrics import get_pascal_voc_metrics, MetricPerClass, get_bounding_boxes

with open('tests/sample/groundtruths_coco.json') as fp:
    gold_dataset = coco_decoder.load_true_object_detection_dataset(fp)
with open('tests/sample/detections_coco.json') as fp:
    pred_dataset = coco_decoder.load_pred_object_detection_dataset(fp, gold_dataset)

gt_BoundingBoxes = get_bounding_boxes(gold_dataset)
pd_BoundingBoxes = get_bounding_boxes(pred_dataset)
results = get_pascal_voc_metrics(gt_BoundingBoxes, pd_BoundingBoxes, .5)

ap, precision, recall, tp, fp, etc

for cls, metric in results.items():
    label = metric.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.metrics import MetricPerClass
mAP = MetricPerClass.mAP(results)

IoU

from podm.box import Box, intersection_over_union

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

Official COCO Eval

from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval

coco_gld = COCO('tests/sample/groundtruths_coco.json')
coco_rst = coco_gld.loadRes('tests/sample/detections_coco.json')
cocoEval = COCOeval(coco_gld, coco_rst, iouType='bbox')
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()

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
• Official COCO Eval

License

Copyright BioNLP Lab at Weill Cornell Medicine, 2022.

Distributed under the terms of the MIT license, this is free and open source software.