Systems and methods of contrastive point completion with fine-to-coarse refinement

1. A method of recovering a complete and dense point cloud from a partial point cloud, comprising:

constructing a sparse but complete point cloud from the partial point cloud through a contrastive teacher-student neural network; and

transforming the sparse but complete point cloud to the complete and dense point cloud, further including:

obtaining a symmetrical point cloud from the partial point cloud; and

refining shape of the sparse but complete point cloud to get a refined sparse but complete point cloud, including:

predicting offsets to points in the sparse but complete point cloud to match a merged point cloud from the partial point cloud, the symmetrical point cloud, and the sparse but complete point cloud; and

applying the offsets to the points in the sparse but complete point cloud to get the refined sparse but complete point cloud.

2. The method according to claim 1 , wherein the partial point cloud is obtained from a raw scanning.

3. The method according to claim 1 , wherein the contrastive teacher-student neural network has a dual network structure comprising a teacher network and a student network both sharing same architecture, wherein

the teacher network is a point cloud self-reconstruction network that takes a ground-truth complete sparse point cloud as input, and

the student network is a point cloud completion network that takes the partial point cloud as input.

4. The method according to claim 3 , wherein constructing the sparse but complete point cloud from the partial point cloud through the contrastive teacher-student neural network includes:

training the student network from ground-truth data and features learned from the teacher network so that a first feature representation of the partial point cloud of the student network is close to a second feature representation of the complete sparse point cloud of the teacher network,

wherein the teacher network and the student network are trained jointly.

5. The method according to claim 4 , wherein a difference between the first feature representation of the partial point cloud of the student network and the second feature representation of the complete sparse point cloud of the teacher network is minimized after the training.

6. The method according to claim 3 , wherein:

the student network has a first encoder that learns a first feature representation of the partial point cloud, and the teacher network has a second encoder that learns a second feature representation of the complete sparse point cloud, the first encoder and the second encoder sharing same network architecture of PointNet with different weights; and

the student network has a first decoder that predicts the first feature representation of the partial point cloud, and the teacher network has a second decoder that predicts the second feature representation of the complete sparse point cloud, the first decoder and the second decoder sharing same multilayer perceptron (MLP) architecture with different weights.

7. The method according to claim 1 , wherein transforming the sparse but complete point cloud to the complete and dense point cloud further includes:

iteratively up-sampling the refined sparse but complete point cloud to get an initial complete and dense point cloud.

8. The method according to claim 7 , wherein transforming the sparse but complete point cloud to the complete and dense point cloud further includes: mixing the partial point cloud, the symmetrical point cloud and the initial complete and dense point cloud to form the complete and dense point cloud.

9. The method according to claim 7 , wherein applying the symmetry detection and enrichment includes:

detecting a symmetry plane for the sparse but complete point cloud; and

in accordance with a determination that the symmetry plane exists, applying planar symmetry for points within the partial point cloud to get the symmetrical point cloud,

in accordance with a determination that the symmetry plane does not exist, setting the symmetrical point cloud the same as the partial point cloud.

10. The method according to claim 1 , wherein

obtaining the symmetrical point cloud from the partial point cloud includes applying symmetry detection and enrichment based on the sparse but complete point cloud to get the symmetrical point cloud from the partial point cloud.

11. The method according to claim 7 , wherein iteratively up-sampling includes:

iteratively up-sampling the refined sparse but complete point cloud by a ratio greater than 1; and

de-noising an up-sampled point cloud at each up-sampling iteration except last iteration.

12. The method according to claim 11 , wherein de-noising further includes:

determining a score for each point in the up-sampled point cloud based on a comparison with a ground-truth complete point cloud, and

dropping a point in the up-sampled point cloud if a respective score is exceeding a predetermined threshold.

13. An electronic apparatus comprising one or more processing units, memory coupled to the one or more processing units, and a plurality of programs stored in the memory that, when executed by the one or more processing units, cause the electronic apparatus to perform a plurality of operations of recovering a complete and dense point cloud from a partial point cloud, comprising:

constructing a sparse but complete point cloud from the partial point cloud through a contrastive teacher-student neural network; and

transforming the sparse but complete point cloud to the complete and dense point cloud, further including:

obtaining a symmetrical point cloud from the partial point cloud; and

refining shape of the sparse but complete point cloud to get a refined sparse but complete point cloud, including:

predicting offsets to points in the sparse but complete point cloud to match a merged point cloud from the partial point cloud, the symmetrical point cloud, and the sparse but complete point cloud; and

applying the offsets to the points in the sparse but complete point cloud to get the refined sparse but complete point cloud.

14. The electronic apparatus according to claim 13 , wherein the contrastive teacher-student neural network has a dual network structure comprising a teacher network and a student network both sharing same architecture, wherein

the teacher network is a point cloud self-reconstruction network that takes a ground-truth complete sparse point cloud as input, and

the student network is a point cloud completion network that takes the partial point cloud as input.

15. The electronic apparatus according to claim 14 , wherein constructing the sparse but complete point cloud from the partial point cloud through the contrastive teacher-student neural network includes:

training the student network from ground-truth data and features learned from the teacher network so that a first feature representation of the partial point cloud of the student network is close to a second feature representation of the complete sparse point cloud of the teacher network,

wherein the teacher network and the student network are trained jointly.

16. The electronic apparatus according to claim 15 , wherein a difference between the first feature representation of the partial point cloud of the student network and the second feature representation of the complete sparse point cloud of the teacher network is minimized after the training.

17. The electronic apparatus according to claim 13 , wherein transforming the sparse but complete point cloud to the complete and dense point cloud further includes:

iteratively up-sampling the refined sparse but complete point cloud to get an initial complete and dense point cloud;

wherein obtaining the symmetrical point cloud from the partial point cloud includes applying symmetry detection and enrichment based on the sparse but complete point cloud to get the symmetrical point cloud from the partial point cloud.

18. The electronic apparatus according to claim 17 , wherein transforming the sparse but complete point cloud to the complete and dense point cloud further includes: mixing the partial point cloud, the symmetrical point cloud and the initial complete and dense point cloud to form the complete and dense point cloud.

19. A non-transitory computer readable storage medium storing a plurality of programs for execution by an electronic apparatus having one or more processing units, wherein the plurality of programs, when executed by the one or more processing units, cause the electronic apparatus to perform a plurality of operations of recovering a complete and dense point cloud from a partial point cloud, comprising:

constructing a sparse but complete point cloud from the partial point cloud through a contrastive teacher-student neural network; and

transforming the sparse but complete point cloud to the complete and dense point cloud, further including:

obtaining a symmetrical point cloud from the partial point cloud; and

refining shape of the sparse but complete point cloud to get a refined sparse but complete point cloud, including:

predicting offsets to points in the sparse but complete point cloud to match a merged point cloud from the partial point cloud, the symmetrical point cloud, and the sparse but complete point cloud; and

applying the offsets to the points in the sparse but complete point cloud to get the refined sparse but complete point cloud.

20. The non-transitory computer readable storage medium according to claim 19 , wherein the contrastive teacher-student neural network has a dual network structure comprising a teacher network and a student network both sharing same architecture, wherein

the teacher network is a point cloud self-reconstruction network that takes a ground-truth complete sparse point cloud as input, and

the student network is a point cloud completion network that takes the partial point cloud as input.