Limits...
Forest Walk Methods for Localizing Body Joints from Single Depth Image.

Jung HY, Lee S, Heo YS, Yun ID - PLoS ONE (2015)

Bottom Line: A regression forest is trained to estimate the probability distribution to the direction or offset toward the particular joint, relative to the adjacent position.The distribution for next position is found from traversing the regression tree from new position.The continual position sampling through 3D space will eventually produce an expectation of sample positions, which we estimate as the joint position.

View Article: PubMed Central - PubMed

Affiliation: Division of Computer and Electronic Systems Engineering, Hankuk University of Foreign Studies, Yongin, Republic of Korea.

ABSTRACT
We present multiple random forest methods for human pose estimation from single depth images that can operate in very high frame rate. We introduce four algorithms: random forest walk, greedy forest walk, random forest jumps, and greedy forest jumps. The proposed approaches can accurately infer the 3D positions of body joints without additional information such as temporal prior. A regression forest is trained to estimate the probability distribution to the direction or offset toward the particular joint, relative to the adjacent position. During pose estimation, the new position is chosen from a set of representative directions or offsets. The distribution for next position is found from traversing the regression tree from new position. The continual position sampling through 3D space will eventually produce an expectation of sample positions, which we estimate as the joint position. The experiments show that the accuracy is higher than current state-of-the-art pose estimation methods with additional advantage in computation time.

No MeSH data available.


Related in: MedlinePlus

The offset positions are randomly sampled from head and chest joints.(a) illustrates offset sample range spheres in green. In (b), the green dots represents offset samples.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4581738&req=5

pone.0138328.g003: The offset positions are randomly sampled from head and chest joints.(a) illustrates offset sample range spheres in green. In (b), the green dots represents offset samples.

Mentions: Compared to the previous random forest pose estimation methods, our approach trains a separate regression for each body joint, and the objective function is a simple sum of 3D squared errors. In Girshick et. al’s paper [10], a single regression tree is trained to learn offsets to all body parts. This leads to very large memory requirement at the leaves, and a very high dimensioned error sum to be minimized. They employ vote compression at the leaves. Also, per joint distance thresholds are used to eliminate joint offsets that are too far away. In our approach, each tree is specialized for each body joint, which allows for different and appropriate training set for each joint. For example, in order to train for head, we only need samples around the head as exemplified in Fig 3. Points around feet are not included in the training set for the head because features around feet will not likely tell you where the head is. Like this, we are able to naturally eliminate irrelevant samples from the training set. In the previous approach, all offsets to body joints are trained by a single tree, which is not an effective and efficient approach [10] without the additional per joint distance thresholds and vote compressions. We were able to circumvent the two major problems addressed by [10], by finding a way to train specific regression tree for each joint.


Forest Walk Methods for Localizing Body Joints from Single Depth Image.

Jung HY, Lee S, Heo YS, Yun ID - PLoS ONE (2015)

The offset positions are randomly sampled from head and chest joints.(a) illustrates offset sample range spheres in green. In (b), the green dots represents offset samples.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4581738&req=5

pone.0138328.g003: The offset positions are randomly sampled from head and chest joints.(a) illustrates offset sample range spheres in green. In (b), the green dots represents offset samples.
Mentions: Compared to the previous random forest pose estimation methods, our approach trains a separate regression for each body joint, and the objective function is a simple sum of 3D squared errors. In Girshick et. al’s paper [10], a single regression tree is trained to learn offsets to all body parts. This leads to very large memory requirement at the leaves, and a very high dimensioned error sum to be minimized. They employ vote compression at the leaves. Also, per joint distance thresholds are used to eliminate joint offsets that are too far away. In our approach, each tree is specialized for each body joint, which allows for different and appropriate training set for each joint. For example, in order to train for head, we only need samples around the head as exemplified in Fig 3. Points around feet are not included in the training set for the head because features around feet will not likely tell you where the head is. Like this, we are able to naturally eliminate irrelevant samples from the training set. In the previous approach, all offsets to body joints are trained by a single tree, which is not an effective and efficient approach [10] without the additional per joint distance thresholds and vote compressions. We were able to circumvent the two major problems addressed by [10], by finding a way to train specific regression tree for each joint.

Bottom Line: A regression forest is trained to estimate the probability distribution to the direction or offset toward the particular joint, relative to the adjacent position.The distribution for next position is found from traversing the regression tree from new position.The continual position sampling through 3D space will eventually produce an expectation of sample positions, which we estimate as the joint position.

View Article: PubMed Central - PubMed

Affiliation: Division of Computer and Electronic Systems Engineering, Hankuk University of Foreign Studies, Yongin, Republic of Korea.

ABSTRACT
We present multiple random forest methods for human pose estimation from single depth images that can operate in very high frame rate. We introduce four algorithms: random forest walk, greedy forest walk, random forest jumps, and greedy forest jumps. The proposed approaches can accurately infer the 3D positions of body joints without additional information such as temporal prior. A regression forest is trained to estimate the probability distribution to the direction or offset toward the particular joint, relative to the adjacent position. During pose estimation, the new position is chosen from a set of representative directions or offsets. The distribution for next position is found from traversing the regression tree from new position. The continual position sampling through 3D space will eventually produce an expectation of sample positions, which we estimate as the joint position. The experiments show that the accuracy is higher than current state-of-the-art pose estimation methods with additional advantage in computation time.

No MeSH data available.


Related in: MedlinePlus