ml.utils — ML Utils

The ml.utils module contains utility classes and functions for machine learning.

class NMS - Soft-Non-Maximum Suppression

The NMS object is used to collect a list of bounding boxes and their associated scores and then filter out overlapping bounding boxes with lower scores. Additionally, it remaps bounding boxes detected in a sub-window back to the original image coordinates.

Constructors

class ml.utils.NMS(window_w: int, window_h: int, roi: tuple[int, int, int, int]) → NMS

Creates a NMS object with the given window size and region of interest (ROI). The window is width/height of the input tensor of image model. The ROI is the region of interest that returned by the Normalization() object which corresponds to the region of the image that the model was run on. This allows the NMS object to remap bounding boxes detected in a sub-window back to the original image coordinates.

Methods

add_bounding_boxes(xmin: float, ymin: float, xmax: float, ymax: float, score: float, label_index: int) → None

Adds a bounding box to the NMS object with the given coordinates, score, and label index.

xmin, ymin, xmax, and ymax are the bounding box coordinates in the range of 0.0 to 1.0 where (0.0, 0.0) is the top-left corner of the image and (1.0, 1.0) is the bottom-right corner of the image.

score is the confidence score of the bounding box (0.0-1.0).

label_index is the index of the label associated with the bounding box.

get_bounding_boxes(threshold: float = 0.1, sigma: float = 0.1) → list[tuple[int, int, int, int, float, int]]

Returns a list of bounding boxes that have been filtered by the NMS object and remapped to the original image coordinates. Bounding box tuples are (x, y, w, h, score, label_index). After calling this method you should create a new NMS object if you want to process a new set of bounding boxes.

Bounding boxes must have a higher score then threshold to be kept.

sigma controls the gaussian used to apply a score penalty to overlapping bounding boxes using the Soft-Non-Maximum-Suppression algorithm. A higher sigma will result in a more aggressive suppression of overlapping bounding boxes.