Skip to content
EvalSuite
Documentation menu

Object detection

Average precision

Implementeddetection.average_precision

Definition

Area under the interpolated precision-recall curve for one class at one IoU threshold.

Formula

AP = Σ (r_{n+1} − r_n) · p_interp(r_{n+1})

Range: [0, 1]

Inputs and outputs

  • ground truth: boxes [x1, y1, x2, y2] with class labels
  • predictions: boxes with class labels and confidence scores
  • iou_threshold

Returns: float per class

Assumptions

No assumptions beyond valid, aligned inputs of the documented types.

Limitations

  • Interpolation and matching conventions differ between VOC and COCO; the convention used is documented with every result.

Python API

PythonSince v0.3.0
import evalsuite as es

es.average_precision_detection(gt, preds, iou_threshold=0.5)  # per class
es.average_precision_detection(gt, preds, interpolation="voc")

References

  1. Everingham, M., Van Gool, L., Williams, C. K. I., Winn, J., & Zisserman, A. (2010). The PASCAL Visual Object Classes (VOC) challenge. International Journal of Computer Vision, 88(2), 303–338.

Implementation status