The goal of this 3D-mesh segmentation benchmark is to provide an automatic tool to evaluate, analyse, and compare the different automatic 3D-mesh segmentation algorithms.
The corpus of the benchmark contains 3D-models (as triangle meshes) grouped in classes. Each 3D-model of the corpus is associated with some manual segmentations done by volunteers. We have selected a small number of varied models with respect to a set of properties. All the selected models are manifold, connected, and do not have intersecting faces. Hence they are supported as an input by any segmentation algorithm. In order to collect precise manual segmentations, we have assisted the volunteers in tracing the vertex-boundaries through the different models. Note that the volunteers have freely segmented the models and no condition was imposed on the manner with which they have segmented them.
The repository contains twenty-eight 3D-models grouped in five classes, namely animal, furniture, hand, human and bust. Each 3D-model of the corpus is associated with 4 manual segmentations which give a total of 112 ground-truth segmentations done by 36 volunteers.
In order to evaluate your segmentation algorithm you have to create a user account. Then, upload the lab files which are the results of applying your segmentation algorithm on the corpus models. Finally, the evaluation result will be displayed in a temporary page. For more details please go to the help page.
The Results page contains evaluation results of different methods. The results are illustrated in plot format. You can visualize one or multiple methods in the plot. The mean score of each corpus-class together with the global-mean are displayed in a table for each method. You can export the evaluation results (Plot, mean-score, etc.) under different formats (PDF, CVS, etc.).
This 3D Segmentation Benchmark is done within the framework of the MADRAS project (3D Models And Dynamic models Representation And Segmentation - ref. ANR-07-MDCO-015) supported by the French Government Research Department, from January 2008 to September 2011.
If you use any part of this benchmark, please cite:
Halim Benhabiles, Jean-Philippe Vandeborre, Guillaume Lavoué and Mohamed Daoudi "A framework for the objective evaluation of segmentation algorithms using a ground-truth of human segmented 3D-models" IEEE International Conference on Shape Modeling and Applications (SMI), Beijing, China, June 26-28, 2009.