Functionnals

Here are functions that are used inside Predictor class and metrics.

Metrics

Here are all the functions that compose the processes above.

detectools.metrics.functionnals.match_boxes(prediction: Format, target: Format, iou_threshold: float = 0.5) → Tuple[Tensor, Tensor, Tensor, Tuple[Tensor, Tensor]][source]

Match better prediction boxes candidates with target boxes. Return indexes of prediction and target boxes that match and compute statistics of detection quality (Tp, FP, FN).

Parameters:

  • prediction (Format) – Prediction.

  • target (Format) – Target.

  • iou_threshold (float, optional) – IoU threshold to discard some matchs with overlapping < to thr. Defaults to 0.5.

Returns:

Detection statsitics (TP, FN, FN) & prediction and target boxes indexes that match well.

Return type:

Tuple[Tensor, Tensor, Tensor, Tuple[Tensor, Tensor]]

detectools.metrics.functionnals.precision(tp, fp, tn, fn)[source]
detectools.metrics.functionnals.recall(tp, fp, tn, fn)[source]
detectools.metrics.functionnals.f1score(tp, fp, tn, fn)[source]
detectools.metrics.functionnals.iou(tp: Tensor, fp: Tensor, tn: Tensor, fn: Tensor) → Tensor[source]

Compute IoU from statistics.

detectools.metrics.functionnals.accuracy(tp: Tensor, fp: Tensor, tn: Tensor, fn: Tensor)[source]

Compute accuracy from statistics.

Inference

detectools.inference.engine.add_offset(image: Tensor, offset: int) → Tensor[source]

Pad fix number of pixels around sides of image :param image: Image to pad. :type image: Tensor :param offset: Number of pixels to pad. :type offset: int

Returns:

  • Padded image.

Return type:

Tensor

detectools.inference.engine.remove_offset(image: Tensor, offset: int) → Tensor[source]

Remove fix number of pixels on each side of image.

Parameters:

  • image (Tensor) – Image.

  • offset (int) – Border size to remove.

  • image – Image to pad.

  • offset – Number of pixels to remove.

Returns:

Image without offset.

Return type:

Tensor

Returns:

  • Cropped image.

Return type:

Tensor

detectools.inference.engine.pad_to(image: Tensor, size: Tuple, fill_value: int = 0) → Tensor[source]

Pad Image with fill value to fit size with origin image in center of padded image.

Parameters:

  • image (Tensor) – Image to be padded.

  • size (Tuple) – Size to fit with padding.

  • fill_value (int, optional) – value to use for filling new pixels.

  • image – Image to be padded.

  • size – Size to fit with padding.

  • fill_value – Value to use for filling new pixels. Defaults to 0.

Returns:

Padded image

Return type:

Tensor

Returns:

  • Padded Image

Return type:

Tensor

detectools.inference.engine.crop_to(image: Tensor, size: Tuple) → Tensor[source]

Crop Image in center.

Parameters:

  • image (Tensor) – Image to crop.

  • size (Tuple) – Cropped size.

Returns:

  • Cropped Image.

Return type:

Tensor

detectools.inference.engine.patchification(image: Tensor, patch_size: Tuple, overlap: float) → Tuple[Tensor, List[Tuple[int]], Tuple[int]][source]

Cut image in patches according to patch size and overlapping. If needed padding is applied to fit patch size multiplicator on H & W.

Parameters:

  • image (Tensor) – Large image to patchify.

  • patch_size (Tuple) – Size of patch.

  • overlap (float) – Overlap between patches.

Returns:

  • Tensor of N patches (N, patch_height, patch_width)

  • Coordinates of patches.

Return type:

Tuple[Tensor, List[Tuple[int]], Tuple[int]]

detectools.inference.engine.unpatchification(predictions: List[BaseFormat], coordinates: List[Tuple[int]], spatial_size: Tuple[int]) → BaseFormat[source]

Build a prediction from multiple patch predictions and coordinates.

Parameters:

  • predictions (List[BaseFormat]) – Patches predictions.

  • coordinates (List[Tuple[int]]) – List of Y, X coordinates.

  • spatial_size (Tuple[int]) – Size of merged prediction.

Returns:

  • BaseFormat of whole image.

Return type:

BaseFormat