`ml.postprocessing.darknet` — Darknet

The ml.postprocessing.darknet module contains post-processors for Darknet based object detection models.

class YoloV2 – YOLO V2

Post-processor for YOLO V2 model outputs.

class ml.postprocessing.darknet.YoloV2(threshold: float = 0.6, anchors: numpy.ndarray = None, nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a YOLO V2 postprocessor.

threshold Score threshold applied before non-maximum suppression.

anchors 2D numpy.ndarray of shape (N, 2) holding (w, h) anchor box dimensions the model was trained on. If None, a built-in default set of 5 anchors is used.

nms_threshold Threshold passed to non-maximum suppression.

nms_sigma Sigma value passed to non-maximum suppression.

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

Invoked by ml.Model.predict() to post-process model outputs.

model The ml.Model instance the post-processor is attached to.

inputs List of model input objects (used to obtain the input ROI).

outputs List of raw model output tensors.

Returns a list of per-class lists of ((x, y, w, h), score) tuples. E.g. [[((x, y, w, h), score), ...], ...]. Empty class lists are kept so each list index matches the model’s class index.

class YoloLC – YOLO LC

Post-processor for YOLO LC model outputs. Subclass of YoloV2 that provides a different default anchor set tuned for the YOLO LC model.

class ml.postprocessing.darknet.YoloLC(threshold: float = 0.6, anchors: numpy.ndarray = None, nms_threshold: float = 0.1, nms_sigma: float = 0.1)

Create a YOLO LC postprocessor.