Method for accelerating three-dimensional object segmentation with point cloud simplifications

1. A device, comprising:

a processing system including a processor; and

a memory that stores executable instructions that, when executed by the processing system, facilitate performance of operations, the operations comprising:

downsampling a full point cloud to obtain a downsampled point cloud, wherein the downsampling reduces a data size of the full point cloud;

using a machine-learning model to assign labels for segmentation and object identification to points in the downsampled point cloud, wherein the machine-learning model is trained on the full point cloud, and wherein the machine-learning model is a submanifold sparse convolutional network; and

selecting model parameters for the machine-learning model, wherein the model parameters comprise a number of feature planes for each layer of a plurality of layers, a style of submanifold block and whether to use optional submanifold blocks.

2. The device of claim 1 , wherein the downsampling uses grid simplification.

3. The device of claim 2 , wherein the downsampled point cloud comprises 20 to 60 percent of points in the full point cloud to reduce memory usage and computational overhead.

4. The device of claim 2 , wherein the downsampled point cloud comprises 63 to 87 percent of points in the full point cloud to improve segmentation accuracy.

5. The device of claim 1 , wherein the operations further comprise selecting a downsampling technique and a downsampling reduction percentage to reduce memory usage while meeting a required segmentation accuracy.

6. The device of claim 1 , wherein the operations further comprise selecting a downsampling technique and a downsampling reduction percentage to reduce computational overhead while meeting a required segmentation accuracy.

7. The device of claim 1 , wherein the processing system comprises a plurality of processors operating in a distributed computing environment, wherein the labels comprise a brand associated with the object, and wherein the brand is used for targeted advertising.

8. A non-transitory machine-readable medium, comprising executable instructions that, when executed by a processing system including a processor, facilitate performance of operations, the operations comprising:

receiving a machine-learning model trained on a full point cloud, wherein the machine-learning model is a submanifold sparse convolutional network;

selecting model parameters for the machine-learning model, wherein the model parameters comprise a number of feature planes for each layer of a plurality of layers, a style of submanifold block and whether to use optional submanifold blocks;

downsampling the full point cloud to a downsampled point cloud using a downsampling technique; and

segmenting the downsampled point cloud using the machine-learning model having the model parameters to assign labels to points in the downsampled point cloud.

9. The non-transitory machine-readable medium of claim 8 , wherein the downsampling technique is grid simplification.

10. The non-transitory machine-readable medium of claim 9 , wherein the downsampled point cloud comprises 20 to 60 percent of points in the full point cloud to reduce memory usage and computational overhead.

11. The non-transitory machine-readable medium of claim 9 , wherein the downsampled point cloud comprises 63 to 87 percent of points in the full point cloud to improve segmenting accuracy.

12. The non-transitory machine-readable medium of claim 8 , wherein the operations further comprise selecting the downsampling technique and a downsampling reduction percentage to reduce memory usage while meeting a required segmentation accuracy.

13. The non-transitory machine-readable medium of claim 8 , wherein the operations further comprise selecting the downsampling technique and a downsampling reduction percentage to reduce computational overhead while meeting a required segmentation accuracy.

14. The non-transitory machine-readable medium of claim 13 , wherein the model parameters comprise a number of feature planes for each layer, a style of submanifold block and whether to use optional submanifold blocks.

15. The non-transitory machine-readable medium of claim 8 , wherein the processing system comprises a plurality of processors operating in a distributed computing environment.

16. A method, comprising:

downsampling, by a processing system including a processor, a full point cloud to a downsampled point cloud using a downsampling technique; and

segmenting, by the processing system, the downsampled point cloud using a submanifold sparse convolutional network having model parameters to assign labels to points in the downsampled point cloud, wherein the downsampling either reduces a memory usage, reduces a computational overhead, improves accuracy, or a combination thereof, and wherein the model parameters comprise a number of feature planes for each layer of a plurality of layers, a style of submanifold block and whether to use optional submanifold blocks.

17. The method of claim 16 , comprising:

selecting, by the processing system, the model parameters for the submanifold sparse convolutional network, the downsampling technique, and a downsampling reduction percentage to achieve a required segmentation accuracy.

18. The method of claim 16 , wherein a label of the labels pertains to a brand of goods associated with an object represented by the downsampled point cloud.

19. The method of claim 16 , comprising:

applying, by the processing system, parameter quantization to reduce a number of bits associated with the downsampled point cloud; and

applying, by the processing system, network pruning and sharing techniques to reduce network complexity and to address overfitting.

20. The method of claim 16 , wherein the downsampling is applied in stages, the method further comprising:

receiving, by the processing system, a user input that identifies regions of the downsampled point cloud that should receive more detail in a first stage of the stages; and

enhancing, by the processing system and based on the receiving of the user input, detail of the regions in a second stage of the stages.