In this paper we analyze the performance of a pipeline for the extraction and semantic labelling of geometrically salient points on acquired human body models, improving the quality of the results and discussing its robustness against pose and body type variations. Following the approach introduced in [10], we use a heat diffusion approach automatically detecting points as maxima of the autodiffusion function and using supervised classification to assign them a semantic label related to the anatomical part where the point is located. The resulting map can be used to perform measurements or to detect pose. The motivation of this approach is related to the fact that landmarks used in traditional anthropometry are not easily identified in digital models being localized by palpation so that an anatomical measurement system should be instead based on geometrical landmarks and it is therefore interesting to find points of this kind that can be recognized in different subjects. The use of heat diffusion analysis enables a robust salient point detection at different levels of detail and the creation of rich point descriptors not depending only on local geometry but also on global context, invariant with respect to articulated deformations (pose variation) and sufficiently stable against changes in body type. In this work we improved the original semantic labelling by selecting optimal point descriptors and feature-space distances and applying a hierarchical coarse to fine approach performing the classification at different scales and propagating the assigned labels from the coarser scale to the finer ones. Furthermore we tested the method on a collection of models representing different body types and with approximately fixed or largely different poses in order to verify the robustness of the semantic labelling. Experimental results show that this approach can be used to recognize robustly at least a selection of landmarks on subjects with different body types and independently on pose and could therefore applied for automatic anthropometric analysis.
Robust Automatic Labelling of Anatomical Landmarks on 3D Body Scans
GIACHETTI, Andrea;CASTELLANI, Umberto;LOVATO, Christian;ZANCANARO, Carlo
2012-01-01
Abstract
In this paper we analyze the performance of a pipeline for the extraction and semantic labelling of geometrically salient points on acquired human body models, improving the quality of the results and discussing its robustness against pose and body type variations. Following the approach introduced in [10], we use a heat diffusion approach automatically detecting points as maxima of the autodiffusion function and using supervised classification to assign them a semantic label related to the anatomical part where the point is located. The resulting map can be used to perform measurements or to detect pose. The motivation of this approach is related to the fact that landmarks used in traditional anthropometry are not easily identified in digital models being localized by palpation so that an anatomical measurement system should be instead based on geometrical landmarks and it is therefore interesting to find points of this kind that can be recognized in different subjects. The use of heat diffusion analysis enables a robust salient point detection at different levels of detail and the creation of rich point descriptors not depending only on local geometry but also on global context, invariant with respect to articulated deformations (pose variation) and sufficiently stable against changes in body type. In this work we improved the original semantic labelling by selecting optimal point descriptors and feature-space distances and applying a hierarchical coarse to fine approach performing the classification at different scales and propagating the assigned labels from the coarser scale to the finer ones. Furthermore we tested the method on a collection of models representing different body types and with approximately fixed or largely different poses in order to verify the robustness of the semantic labelling. Experimental results show that this approach can be used to recognize robustly at least a selection of landmarks on subjects with different body types and independently on pose and could therefore applied for automatic anthropometric analysis.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
