System and method for proposal-free and cluster-free panoptic segmentation system of point clouds

1 . A computer-implemented method for panoptic segmentation of a point cloud, comprising:

projecting a point cloud into a range image;

extracting features from the range image and generating a feature map from the extracted features;

downsampling the feature map and scaling the features during downsampling using local geometry;

extracting features from the downsampled feature map;

semantically segmenting the point cloud at least partially based on the features extracted; and

segmenting instances in the point cloud at least partially based on the features extracted.

2 . The computer-implemented method of claim 1 , wherein the downsampling and the extracting features from the downsampled feature map are repeated.

3 . The computer-implemented method of any of the previous claims , wherein the extracted features are grouped into channels.

4 . The computer-implemented method of claim 3 , further comprising:

weighing the features in each of the channels differently for semantic segmentation and instance segmentation.

5 . The computer-implemented method of claim 1 , wherein points in the point cloud are assigned semantic labels and instance labels during the semantic segmenting and the instance segmenting, respectively, and wherein the semantic labels and the instance labels are projected onto the point cloud.

6 . The computer-implemented method of claim 1 , further comprising:

upsampling the downsampled feature map to a desired resolution; and

processing combined features of the feature map and the downsampled feature map at the desired resolution before the semantically segmenting and the segmenting instances.

7 . The computer-implemented method of claim 6 , wherein the desired resolution is the resolution of the feature map.

8 . The computer-implemented method of claim 6 , wherein the features of the feature map and the downsampled feature map are assigned to feature channels.

9 . The computer-implemented method of claim 8 , wherein the processing includes:

semantically segmenting the point cloud using a first set of weights for the feature channels; and

segmenting instances in the point cloud using a second set of weights for the feature channels, the second set of weights differing from the first set of weights.

10 . The computer-implemented method of claim 1 , further comprising:

dividing foreground points in the point cloud into discrete spaces;

determining a center of mass for each of the discrete spaces based on a position of the foreground points in the discrete space;

determining a distance between the centers of mass of pairs of discrete spaces;

calculating a score for each of the pairs of discrete spaces at least partially based on the distance between the centers of mass of the discrete spaces in the pair; and

grouping the foreground points in each discrete space into objects based on the score for each of the pairs of discrete spaces.

11 . The computer-implemented method of claim 10 , further comprising:

determining the position for each foreground point in the point cloud based on a shifted position from a centroid of an instance to which the point belongs.

12 . The computer-implemented method of claim 10 , wherein the discrete spaces are voxels.

13 . The computer-implemented method of claim 12 , wherein the voxels are unlimited in length along a dimension.

14 . The computer-implemented method of claim 10 , further comprising:

constructing a pairwise distance matrix wherein each matrix element represents the distance between one of the pairs of discrete spaces.

15 . The computer-implemented method of claim 10 , wherein the score, D ij , is calculated as:

D

i

⁢

j

=

e

-

a

⁢

❘

"\[LeftBracketingBar]"

❘

"\[LeftBracketingBar]"

C

D

,

i

-

C

D

,

j

❘

"\[RightBracketingBar]"

❘

"\[RightBracketingBar]"

2

,

where ∥C D, i −C D,j ∥ 2 is the distance between the discrete spaces C D, i and C D, j , and a is a tunable hyperparameter.

16 . A computing system for panoptic segmentation of a point cloud, the computing system comprising:

a processor;

a memory storing machine-executable instructions that, when executed by the processor, cause the processor to:

project a point cloud into a range image;

extract features from the range image and generating a feature map from the extracted features;

downsample the feature map and scaling the features during downsampling using local geometry;

extract features from the downsampled feature map;

semantically segment the point cloud at least partially based on the features extracted; and

segment instances in the point cloud at least partially based on the features extracted.

17 . The computing system of claim 16 , wherein the machine-executable instructions, when executed by the processor, cause the processor to repeat the downsampling and the extraction of features from the downsampled feature map.

18 . The computing system of claim 16 , wherein the extracted features are grouped into channels, and wherein the machine-executable instructions, when executed by the processor, cause the processor to:

weigh the features in each of the channels differently for semantic segmentation and instance segmentation.

19 . The computing system of claim 16 , wherein points in the point cloud are assigned semantic labels and instance labels during the semantic segmenting and the instance segmenting, respectively, and wherein the semantic labels and the instance labels are projected onto the point cloud.

20 . The computing system of claim 16 , wherein the machine-executable instructions, when executed by the processor, cause the processor to:

upsample the downsampled feature map to a desired resolution; and

process combined features of the feature map and the downsampled feature map at the desired resolution before the semantically segmenting and the segmenting instances.