powerboxes 0.2.1

Utility functions to manipulate and compute metrics on boxes

PowerBoxes

Powerboxes is a package containing utility functions for transforming bounding boxes and computing metrics. It is implemented in both Python and Rust. It shows a significant speedup over the equivalent numpy implementations in Python, or other libraries such as shapely or torchvision.

Installation

pip install powerboxes

Python Usage

import powerboxes as pb
import numpy as np

# Create a bounding box
box = np.array([[0, 0, 1, 1]])

# Compute the area of the box
area = pb.box_areas(box)

# Compute the intersection of the box with itself
intersection = pb.iou_distance(box, box)

Functions available

Box Transformations and utilities

• box_areas: Compute the area of list of boxes
• box_convert: Convert a box from one format to another. Supported formats are xyxy, xywh, cxcywh.
• remove_small_boxes: Remove boxes with area smaller than a threshold
• mask_to_boxes: Convert a mask to a list of boxes

Box Metrics

• iou_distance: Compute the intersection over union matrix of two sets of boxes
• parallel_iou_distance: Compute the intersection over union matrix of two sets of boxes in parallel
• giou_distance: Compute the generalized intersection over union matrix of two sets of boxes
• parallel_giou_distance: Compute the generalized intersection over union matrix of two sets of boxes in parallel
• tiou_distance: Compute the tracking intersection over union matrix of two sets of boxes

Rotated Box Metrics

• rotated_iou_distance: Compute the intersection over union matrix of two sets of rotated boxes in cxcywha format
• rotated_giou_distance: Compute the generalized intersection over union matrix of two sets of rotated boxes in cxcywha format

Box NMS

• nms: Non-maximum suppression, returns the indices of the boxes to keep
• rtree_nms: Non-maximum suppression, returns the indices of the boxes to keep, uses a r-tree internally to avoid quadratic complexity, useful when having many boxes.

Benchmarks

Some benchmarks of powerboxes against various open source alternatives, not all functions are benchmarked. Notice that we use log scales, all differences are major !

Box area

Here it's torchvision vs powerboxes vs numpy

Box convert

Here it's torchvision vs powerboxes

Box IoU matrix

Torchvision vs numpy vs powerboxes

NMS

Torchvision vs powerboxes vs lsnms vs numpy

Large image (10000x10000 pixels)

Normal image (1000x1000 pixels)