Estimator training method and pose estimating method using depth image

Patent metadata

Field	Value
Publication number	US-9311713-B2
Application number	US-201414283646-A
Country	US
Kind code	B2
Filing date	May 21, 2014
Priority date	Jun 5, 2013
Publication date	Apr 12, 2016
Grant date	Apr 12, 2016

Abstract

An estimator training method and a pose estimating method using a depth image are disclosed, in which the estimator training method may train an estimator configured to estimate a pose of an object, based on an association between synthetic data and real data, and the pose estimating method may estimate the pose of the object using the trained estimator.

First claim

What is claimed is:

1. A method of training an estimator to estimate a pose of an object, the method comprising:
obtaining a latent tree model of the object; and
generating, by way of a processor, a vector pair configured to partition input data received at a latent node into two parts, with respect to a plurality of latent nodes positioned along paths from a root node to a plurality of leaf nodes of the latent tree model, wherein two vectors included in the vector pair are configured to have directions from a center of the input data toward centers of the partitioned two parts.

2. The method of claim 1 , wherein the latent tree model comprises a binary tree configured to branch from a center of the object to a plurality of predetermined parts of the object based on a joint structure and a kinematic constraint of the object.

3. The method of claim 1 , wherein the generating of the vector pair comprises:
generating, at random, a plurality of candidate vector pairs based on a center of the input data; and
selecting a single candidate vector pair that minimizes a variance of candidate vectors in a subset classified by the corresponding candidate vector pair, from among the plurality of candidate vector pairs.

4. The method of claim 3 , wherein the generating of the vector pair further comprises iteratively performing the generating at random of the plurality of candidate vector pairs and the selecting until a variance of candidate vectors included in the selected candidate vector pair is less than a threshold value.

5. The method of claim 1 , further comprising:
partitioning a test set into N subsets, N being an integer greater than or equal to “2”,
wherein the generating of the vector pair comprises:
calculating two candidate vectors configured to partition one of the partitioned subsets into two parts, based on the corresponding subset;
calculating two offset vectors configured to correct errors of the candidate vectors, based on another of the partitioned subsets; and
calculating two vectors included in the vector pair based on the two candidate vectors and the two offset vectors.

6. The method of claim 1 , wherein the generating comprises:
generating a split node configured to test a partition of the input data; and
generating a division node configured to partition the input data, based on an output of the split node.

7. The method of claim 1 , wherein the generating of the vector pair comprises:
generating a split node configured to test a partition of the input data;
generating an error regression node configured to regress an error of the split node; and
generating a division node configured to partition the input data, based on outputs of the split node and the error regression node.

8. The method of claim 1 , wherein the vector pair comprises candidate vectors each having a variance less than or equal to a predetermined threshold value in a subset classified by the vector pair.

9. The method of claim 1 , wherein the input data comprises at least a portion of a depth image in which the object is photographed.

10. The method of claim 1 , wherein the root node corresponds to a center of the object, and the plurality of leaf nodes corresponds to the plurality of predetermined parts of the object.

11. A method of estimating a pose, the method comprising:
detecting a plurality of parts of an object included in an input image using an estimator trained based on a latent tree model of the object; and
estimating, by way of a processor, a pose of the object based on the detected plurality of parts, wherein the estimator is trained by generating a vector pair configured to partition input data received at a latent node into two parts, with respect to a plurality of latent nodes positioned along paths from a root node to a plurality of leaf nodes of the latent tree model, wherein two vectors included in the vector pair are configured to have directions from a center of the input data toward centers of the partitioned two parts.

12. The method of claim 11 , wherein the estimator is configured to receive the input image at a root node included in the latent tree model, partition the input image into two parts at a plurality of latent nodes included in the latent tree model, and output partitioned images corresponding to the plurality of parts at a plurality of leaf nodes included in the latent tree model.

13. The method of claim 11 , wherein the latent tree model comprises a binary tree configured to branch from a center of the object to a plurality of predetermined parts of the object based on a joint structure and a kinematic constraint of the object.

14. The method of claim 11 , wherein the estimating comprises:
calculating locations of the detected plurality of parts; and
estimating the pose of the object based on the calculated locations.

15. A non-transitory computer-readable storage medium encoded with computer readable code comprising a program for implementing the method of claim 1 .

16. A non-transitory computer-readable storage medium encoded with computer readable code comprising a program for implementing the method of claim 11 .

Assignees

• Samsung Electronics Co Ltd
• Imp Innovations Ltd

Inventors

• Han Jae Joon
• Tang Danhang
• Kim Tae Kyun
• Han Seung Ju
• Yoo Byung In
• Choi Chang Kyu
• Tejani Alykhan
• Chang Hyung Jin

Classifications

• G06V10/7625

  Hierarchical techniques, i.e. dividing or merging patterns to obtain a tree-like representation; Dendograms · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/2155

  characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling · CPC title

• G06V40/11 Primary

  Hand-related biometrics; Hand pose recognition · CPC title

• G06F18/231

  Hierarchical techniques, i.e. dividing or merging pattern sets so as to obtain a dendrogram · CPC title