Imagine being tasked with designing protective clothing for law enforcement officers. The protective gear needs to be effective at stopping hurled projectiles while not severely restricting the movements of the wearer. To design such armor, it's useful to partition the human body into nearly rigid (lower arm and leg) and highly deformable parts (neck, elbow, and so on) capable of exhibiting independent motions. Tough, rigid plates along with flexible (and more expensive) materials can then be used to protect the rigid and deformable parts. Although it's relatively easy to manually define a coarse partition of the human body, defining finer scaled partitions can be quite tricky. For instance, while it may be clear that lower and upper arms are distinct parts, localizing the lower arm-elbow boundary or deciding whether the left and right breast plates are distinct parts is more challenging. To answer such questions, a more data-driven approach is required.
Figure 1: Human subjects of widely varying body shapes and in a variety of poses To this end, in collaboration with Michael Black’s group , we (Soumya Ghosh and Erik Sudderth) have acquired scans of 78 human subjects of varying body shapes. Scans of each subject span multiple poses and capture a wide range of motions. Overall, our dataset consists of 1700 aligned scans (Figure 1).
Figure 2: Prior distribution over body partitions Using a custom statistical model, we're able to infer both the number and spatial extent of body parts from the deformations observed in the dataset. Our model defines a prior distribution over body partitions. The prior places non-zero probability mass only on partitions that comprise exclusively of contiguous parts (Figure 2), pruning away noisy partitions not useful for our application of designing protective clothing.
The model also specifies a distribution over “typical” deformations of body parts. Deformations exhibited by each part and pose combination are then modeled as samples from this distribution. Conditioned on these observed deformations, we use approximate inference to infer a posterior distribution over highly probable partitions (Figure 3).
Figure 3: High probability partitions discovered by our model Qualitatively, the discovered partitions correspond to our intuitions about the body. In addition to capturing large parts, like the head and limbs, the segmentation segregates distinctly moving smaller regions such as elbows, shoulders, biceps, and triceps. More details and quantitative comparisons can be found in . Finally, should the need ever arise to design armor for centaurs, we stand ready.
- D. Hirshberg, M. Loper, E. Rachlin, and M.J. Black. Co-registration: Simultaneous alignment and modeling of articulated 3D shape. In ECCV, pages 242–255, 2012.
- S. Ghosh, E.B. Sudderth, M. Loper, and M.J. Black. From Deformations to Parts: Motion-based Segmentation of 3D Objects. In NIPS, pages 2006-2014, 2012.