Mean-Average-Precision (mAP)

Module Interface

class torchmetrics.detection.mean_ap.MeanAveragePrecision(box_format='xyxy', iou_type='bbox', iou_thresholds=None, rec_thresholds=None, max_detection_thresholds=None, class_metrics=False, **kwargs)

Compute the Mean-Average-Precision (mAP) and Mean-Average-Recall (mAR) for object detection predictions.

\[\text{mAP} = \frac{1}{n} \sum_{i=1}^{n} AP_i\]

where \(AP_i\) is the average precision for class \(i\) and \(n\) is the number of classes. The average precision is defined as the area under the precision-recall curve. For object detection the recall and precision are defined based on the intersection of union (IoU) between the predicted bounding boxes and the ground truth bounding boxes e.g. if two boxes have an IoU > t (with t being some threshold) they are considered a match and therefore considered a true positive. The precision is then defined as the number of true positives divided by the number of all detected boxes and the recall is defined as the number of true positives divided by the number of all ground boxes.

As input to forward and update the metric accepts the following input:

• preds (List): A list consisting of dictionaries each containing the key-values (each dictionary corresponds to a single image). Parameters that should be provided per dict

  • boxes: (FloatTensor) of shape (num_boxes, 4) containing num_boxes detection boxes of the format specified in the constructor. By default, this method expects (xmin, ymin, xmax, ymax) in absolute image coordinates.

  • scores: FloatTensor of shape (num_boxes) containing detection scores for the boxes.

  • labels: IntTensor of shape (num_boxes) containing 0-indexed detection classes for the boxes.

  • masks: bool of shape (num_boxes, image_height, image_width) containing boolean masks. Only required when iou_type=”segm”.

• target (List) A list consisting of dictionaries each containing the key-values (each dictionary corresponds to a single image). Parameters that should be provided per dict:

  • boxes: FloatTensor of shape (num_boxes, 4) containing num_boxes ground truth boxes of the format specified in the constructor. By default, this method expects (xmin, ymin, xmax, ymax) in absolute image coordinates.

  • labels: IntTensor of shape (num_boxes) containing 0-indexed ground truth classes for the boxes.

  • masks: bool of shape (num_boxes, image_height, image_width) containing boolean masks. Only required when iou_type=”segm”.

  • iscrowd: IntTensor of shape (num_boxes) containing 0/1 values indicating whether the bounding box/masks indicate a crowd of objects. Value is optional, and if not provided it will automatically be set to 0.

  • area: FloatTensor of shape (num_boxes) containing the area of the object. Value if optional, and if not provided will be automatically calculated based on the bounding box/masks provided. Only affects which samples contribute to the map_small, map_medium, map_large values

As output of forward and compute the metric returns the following output:

• map_dict: A dictionary containing the following key-values:

  • map: (Tensor), global mean average precision

  • map_small: (Tensor), mean average precision for small objects

  • map_medium:(Tensor), mean average precision for medium objects

  • map_large: (Tensor), mean average precision for large objects

  • mar_1: (Tensor), mean average recall for 1 detection per image

  • mar_10: (Tensor), mean average recall for 10 detections per image

  • mar_100: (Tensor), mean average recall for 100 detections per image

  • mar_small: (Tensor), mean average recall for small objects

  • mar_medium: (Tensor), mean average recall for medium objects

  • mar_large: (Tensor), mean average recall for large objects

  • map_50: (Tensor) (-1 if 0.5 not in the list of iou thresholds), mean average precision at IoU=0.50

  • map_75: (Tensor) (-1 if 0.75 not in the list of iou thresholds), mean average precision at IoU=0.75

  • map_per_class: (Tensor) (-1 if class metrics are disabled), mean average precision per observed class

  • mar_100_per_class: (Tensor) (-1 if class metrics are disabled), mean average recall for 100 detections per image per observed class

  • classes (Tensor), list of all observed classes

For an example on how to use this metric check the torchmetrics mAP example.

Note

map score is calculated with @[ IoU=self.iou_thresholds | area=all | max_dets=max_detection_thresholds ]. Caution: If the initialization parameters are changed, dictionary keys for mAR can change as well. The default properties are also accessible via fields and will raise an AttributeError if not available.

Note

This metric utilizes the official pycocotools implementation as its backend. This means that the metric requires you to have pycocotools installed. In addition we require torchvision version 0.8.0 or newer. Please install with pip install torchmetrics[detection].

Parameters:

• box_format (Literal['xyxy', 'xywh', 'cxcywh']) –

  Input format of given boxes. Supported formats are:

  • ’xyxy’: boxes are represented via corners, x1, y1 being top left and x2, y2 being bottom right.

  • ’xywh’ : boxes are represented via corner, width and height, x1, y2 being top left, w, h being width and height. This is the default format used by pycoco and all input formats will be converted to this.

  • ’cxcywh’: boxes are represented via centre, width and height, cx, cy being center of box, w, h being width and height.

• iou_type (Literal['bbox', 'segm']) – Type of input (either masks or bounding-boxes) used for computing IOU. Supported IOU types are ["bbox", "segm"]. If using "segm", masks should be provided in input.

• iou_thresholds (Optional[List[float]]) – IoU thresholds for evaluation. If set to None it corresponds to the stepped range [0.5,...,0.95] with step 0.05. Else provide a list of floats.

• rec_thresholds (Optional[List[float]]) – Recall thresholds for evaluation. If set to None it corresponds to the stepped range [0,...,1] with step 0.01. Else provide a list of floats.

• max_detection_thresholds (Optional[List[int]]) – Thresholds on max detections per image. If set to None will use thresholds [1, 10, 100]. Else, please provide a list of ints.

• class_metrics (bool) – Option to enable per-class metrics for mAP and mAR_100. Has a performance impact.

• kwargs (Any) – Additional keyword arguments, see Advanced metric settings for more info.

Raises:

• ModuleNotFoundError – If pycocotools is not installed

• ModuleNotFoundError – If torchvision is not installed or version installed is lower than 0.8.0

• ValueError – If box_format is not one of "xyxy", "xywh" or "cxcywh"

• ValueError – If iou_type is not one of "bbox" or "segm"

• ValueError – If iou_thresholds is not None or a list of floats

• ValueError – If rec_thresholds is not None or a list of floats

• ValueError – If max_detection_thresholds is not None or a list of ints

• ValueError – If class_metrics is not a boolean

Example

>>> from torch import tensor
>>> from torchmetrics.detection import MeanAveragePrecision
>>> preds = [
...   dict(
...     boxes=tensor([[258.0, 41.0, 606.0, 285.0]]),
...     scores=tensor([0.536]),
...     labels=tensor([0]),
...   )
... ]
>>> target = [
...   dict(
...     boxes=tensor([[214.0, 41.0, 562.0, 285.0]]),
...     labels=tensor([0]),
...   )
... ]
>>> metric = MeanAveragePrecision()
>>> metric.update(preds, target)
>>> from pprint import pprint
>>> pprint(metric.compute())
{'classes': tensor(0, dtype=torch.int32),
 'map': tensor(0.6000),
 'map_50': tensor(1.),
 'map_75': tensor(1.),
 'map_large': tensor(0.6000),
 'map_medium': tensor(-1.),
 'map_per_class': tensor(-1.),
 'map_small': tensor(-1.),
 'mar_1': tensor(0.6000),
 'mar_10': tensor(0.6000),
 'mar_100': tensor(0.6000),
 'mar_100_per_class': tensor(-1.),
 'mar_large': tensor(0.6000),
 'mar_medium': tensor(-1.),
 'mar_small': tensor(-1.)}

static coco_to_tm(coco_preds, coco_target, iou_type='bbox')

Utility function for converting .json coco format files to the input format of this metric.

The function accepts a file for the predictions and a file for the target in coco format and converts them to a list of dictionaries containing the boxes, labels and scores in the input format of this metric.

Parameters:

• coco_preds (str) – Path to the json file containing the predictions in coco format

• coco_target (str) – Path to the json file containing the targets in coco format

• iou_type (Literal['bbox', 'segm']) – Type of input, either bbox for bounding boxes or segm for segmentation masks

Returns:

List of dictionaries containing the predictions in the input format of this metric target: List of dictionaries containing the targets in the input format of this metric

Return type:

preds

Example

>>> # File formats are defined at https://cocodataset.org/#format-data
>>> # Example files can be found at
>>> # https://github.com/cocodataset/cocoapi/tree/master/results
>>> from torchmetrics.detection import MeanAveragePrecision
>>> preds, target = MeanAveragePrecision.coco_to_tm(
...   "instances_val2014_fakebbox100_results.json.json",
...   "val2014_fake_eval_res.txt.json"
...   iou_type="bbox"
... )  

plot(val=None, ax=None)

Plot a single or multiple values from the metric.

Parameters:

• val (Union[Dict[str, Tensor], Sequence[Dict[str, Tensor]], None]) – Either a single result from calling metric.forward or metric.compute or a list of these results. If no value is provided, will automatically call metric.compute and plot that result.

• ax (Optional[Axes]) – An matplotlib axis object. If provided will add plot to that axis

Return type:

Tuple[Figure, Union[Axes, ndarray]]

Returns:

Figure object and Axes object

Raises:

ModuleNotFoundError – If matplotlib is not installed

>>> from torch import tensor
>>> from torchmetrics.detection.mean_ap import MeanAveragePrecision
>>> preds = [dict(
...     boxes=tensor([[258.0, 41.0, 606.0, 285.0]]),
...     scores=tensor([0.536]),
...     labels=tensor([0]),
... )]
>>> target = [dict(
...     boxes=tensor([[214.0, 41.0, 562.0, 285.0]]),
...     labels=tensor([0]),
... )]
>>> metric = MeanAveragePrecision()
>>> metric.update(preds, target)
>>> fig_, ax_ = metric.plot()

>>> # Example plotting multiple values
>>> import torch
>>> from torchmetrics.detection.mean_ap import MeanAveragePrecision
>>> preds = lambda: [dict(
...     boxes=torch.tensor([[258.0, 41.0, 606.0, 285.0]]) + torch.randint(10, (1,4)),
...     scores=torch.tensor([0.536]) + 0.1*torch.rand(1),
...     labels=torch.tensor([0]),
... )]
>>> target = [dict(
...     boxes=torch.tensor([[214.0, 41.0, 562.0, 285.0]]),
...     labels=torch.tensor([0]),
... )]
>>> metric = MeanAveragePrecision()
>>> vals = []
>>> for _ in range(20):
...     vals.append(metric(preds(), target))
>>> fig_, ax_ = metric.plot(vals)

tm_to_coco(name='tm_map_input')

Utility function for converting the input for this metric to coco format and saving it to a json file.

This function should be used after calling .update(…) or .forward(…) on all data that should be written to the file, as the input is then internally cached. The function then converts to information to coco format a writes it to json files.

Parameters:

name (str) – Name of the output file, which will be appended with “_preds.json” and “_target.json”

Return type:

None

Example

>>> from torch import tensor
>>> from torchmetrics.detection import MeanAveragePrecision
>>> preds = [
...   dict(
...     boxes=tensor([[258.0, 41.0, 606.0, 285.0]]),
...     scores=tensor([0.536]),
...     labels=tensor([0]),
...   )
... ]
>>> target = [
...   dict(
...     boxes=tensor([[214.0, 41.0, 562.0, 285.0]]),
...     labels=tensor([0]),
...   )
... ]
>>> metric = MeanAveragePrecision()
>>> metric.update(preds, target)
>>> metric.tm_to_coco("tm_map_input")