kitti

The KITTI 3D object detection protocol: per-class AP at the official IoU thresholds.

Functions:

kitti_average_precision –

The KITTI 3D detection metrics: per-class AP at the official IoU thresholds, and their mean.

kitti_average_precision [source] ¶

kitti_average_precision(
    matches: Sequence[BoxMatches],
    class_names: Sequence[str] = (
        "Car",
        "Pedestrian",
        "Cyclist",
    ),
) -> Dict[str, float]

The KITTI 3D detection metrics: per-class AP at the official IoU thresholds, and their mean.

Each class is matched at its KITTI IoU threshold (Car at \(0.7\), Pedestrian and Cyclist at \(0.5\)), and AP interpolates precision at the 11 recall points \(0.0, 0.1, \ldots, 1.0\) of the KITTI benchmark. mAP averages the per-class APs.

Parameters:

matches (Sequence[BoxMatches]) –

The box_matches records of every evaluated scene, with labels indexing class_names.
class_names (Sequence[str], default: ('Car', 'Pedestrian', 'Cyclist') ) –

Class name per label index, each one of Car, Pedestrian and Cyclist.

Returns:

Dict[str, float] –

AP/<class> for every class, and mAP.

Example

>>> import torch
>>> from torch_pointcloud.metrics import box_matches, kitti_average_precision
>>> zero = torch.tensor([0])
>>> boxes = torch.tensor([[0.0, 0.0, 0.0, 4.0, 2.0, 1.5, 0.0]])
>>> preds = {"boxes": boxes, "scores": torch.tensor([0.9]), "labels": zero, "batch": zero}
>>> target = {"boxes": boxes, "labels": zero, "batch": zero}
>>> metrics = kitti_average_precision([box_matches(preds, target)])
>>> sorted(metrics)
['AP/Car', 'AP/Cyclist', 'AP/Pedestrian', 'mAP']