ml.postprocessing.mediapipe — Google Mediapipe

The ml.postprocessing.mediapipe module contains post-processors for Google Mediapipe models.

class mediapipe_detection_postprocess – Generic Mediapipe Detector

Base class shared by BlazeFace and BlazePalm. Decodes anchor-based bounding boxes and keypoints, then performs NMS.

class ml.postprocessing.mediapipe.mediapipe_detection_postprocess(threshold: float = 0.6, anchors: ndarray | None = None, anchor_grid: list[tuple[int, int]] | None = None, scores: list[int] = [], cords: list[int] = [], nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a generic Mediapipe detection post-processor.

threshold Score threshold applied to raw logits before sigmoid.

anchors Optional pre-built anchor array of shape (N, 2) containing (cx, cy) centers normalized to [0, 1]. If None, anchors are generated from anchor_grid.

anchor_grid List of (grid_size, scales) tuples used to generate anchors when anchors is None.

scores List of model output indices that contain score tensors.

cords List of model output indices that contain box/keypoint tensors.

nms_threshold IoU threshold for non-maximum suppression.

nms_sigma Sigma for soft-NMS score decay.

__call__(model: ml.Model, inputs: list, outputs: list) → list

Run post-processing on model outputs and return a list of ((x, y, w, h), score, keypoints) tuples.

detection_post_process(ih: int, iw: int, nms: ml.utils.NMS, model: ml.Model, inputs: list, outputs: list, score_idx: int, cords_idx: int, t: float, anchors: ndarray) → None

Decode and add bounding boxes from a single (score, cords) output pair into the supplied NMS accumulator.

class BlazeFace – Face Detection

Post-processes BlazeFace model output.

class ml.postprocessing.mediapipe.BlazeFace(threshold: float = 0.6, anchors: ndarray | None = None, nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a BlazeFace post-processor. Uses an anchor grid of [(16, 2), (8, 6)] with score outputs [1, 2] and box outputs [0, 3].

threshold Score threshold for detections.

anchors Optional pre-built anchor array; generated automatically if None.

nms_threshold IoU threshold for non-maximum suppression.

nms_sigma Sigma for soft-NMS score decay.

Returns [((x, y, w, h), score, keypoints)] from __call__, where keypoints is a list of (x, y) points.

class BlazePalm – Palm Detection

Post-processes BlazePalm model output.

class ml.postprocessing.mediapipe.BlazePalm(threshold: float = 0.6, anchors: ndarray | None = None, nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a BlazePalm post-processor. Uses an anchor grid of [(24, 2), (12, 6)] with score outputs [0] and box outputs [1].

threshold Score threshold for detections.

anchors Optional pre-built anchor array; generated automatically if None.

nms_threshold IoU threshold for non-maximum suppression.

nms_sigma Sigma for soft-NMS score decay.

Returns [((x, y, w, h), score, keypoints)] from __call__, where keypoints is a list of (x, y) points.

class FaceLandmarks – Face Landmarks

Post-processes FaceLandmarks model output.

class ml.postprocessing.mediapipe.FaceLandmarks(threshold: float = 0.6, nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a FaceLandmarks post-processor.

threshold Score threshold (after sigmoid) for accepting a detection.

nms_threshold IoU threshold for non-maximum suppression.

nms_sigma Sigma for soft-NMS score decay.

Returns ((x, y, w, h), score, keypoints) from __call__, where keypoints is a list of (x, y, z) points.

class HandLandmarks – Hand Landmarks

Post-processes HandLandmarks model output.

class ml.postprocessing.mediapipe.HandLandmarks(threshold: float = 0.6, nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a HandLandmarks post-processor.

threshold Score threshold for accepting a detection.

nms_threshold IoU threshold for non-maximum suppression.

nms_sigma Sigma for soft-NMS score decay.

Returns [[((x, y, w, h), score, keypoints)]] from __call__, with one inner list per handedness class (left=0, right=1). keypoints is a list of (x, y, z) points. Empty class lists are preserved so each list’s index matches the class index.

class MoveNet – Pose Estimation

Post-processes MoveNet single-pose model output.

class ml.postprocessing.mediapipe.MoveNet(threshold: float = 0.6, nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a MoveNet post-processor.

threshold Per-keypoint confidence threshold; keypoints below this value are excluded from the bounding box and mean score.

nms_threshold IoU threshold for non-maximum suppression.

nms_sigma Sigma for soft-NMS score decay.

Returns ((x, y, w, h), score, keypoints) from __call__, where keypoints is a list of (x, y, score) points in input pixel coordinates.