open3d.ml.tf.ops.nms

open3d.ml.tf.ops.nms(boxes, scores, nms_overlap_thresh, name=None)

Performs non-maximum suppression of bounding boxes.

This function performs non-maximum suppression for the input bounding boxes considering the the per-box score and overlaps. It returns the indices of the selected boxes.

Minimal example:

import open3d.ml.tf as ml3d
import numpy as np

boxes = np.array([[15.0811, -7.9803, 15.6721, -6.8714, 0.5152],
                  [15.1166, -7.9261, 15.7060, -6.8137, 0.6501],
                  [15.1304, -7.8129, 15.7069, -6.8903, 0.7296],
                  [15.2050, -7.8447, 15.8311, -6.7437, 1.0506],
                  [15.1343, -7.8136, 15.7121, -6.8479, 1.0352],
                  [15.0931, -7.9552, 15.6675, -7.0056, 0.5979]],
                 dtype=np.float32)
scores = np.array([3, 1.1, 5, 2, 1, 0], dtype=np.float32)
nms_overlap_thresh = 0.7
keep_indices = ml3d.ops.nms(boxes, scores, nms_overlap_thresh)
print(keep_indices)

# PyTorch example.
import torch
import open3d.ml.torch as ml3d

boxes = torch.Tensor([[15.0811, -7.9803, 15.6721, -6.8714, 0.5152],
                      [15.1166, -7.9261, 15.7060, -6.8137, 0.6501],
                      [15.1304, -7.8129, 15.7069, -6.8903, 0.7296],
                      [15.2050, -7.8447, 15.8311, -6.7437, 1.0506],
                      [15.1343, -7.8136, 15.7121, -6.8479, 1.0352],
                      [15.0931, -7.9552, 15.6675, -7.0056, 0.5979]])
scores = torch.Tensor([3, 1.1, 5, 2, 1, 0])
nms_overlap_thresh = 0.7
keep_indices = ml3d.ops.nms(boxes, scores, nms_overlap_thresh)
print(keep_indices)
Parameters
  • boxes – A Tensor. Must be one of the following types: float32. (N, 5) float32 tensor. Bounding boxes are represented as (x0, y0, x1, y1, rotate).

  • scores – A Tensor. Must have the same type as boxes. (N,) float32 tensor. A higher score means a more confident bounding box.

  • nms_overlap_thresh – A float. float value between 0 and 1. When a high-score box is selected, other remaining boxes with IoU > nms_overlap_thresh will be discarded. A higher nms_overlap_thresh means more boxes will be kept.

  • name – A name for the operation (optional).

Returns

A Tensor of type int64. (M,) int64 tensor. The selected box indices.