Title : 4D Data Acquisition



Project Lead : Edmond Boyer From : INRIA (None)

Dates : from -- to 2014-02-18 16:37:09

Description :

Motivation and objectives :
This project is about the acquisition of 4D sequences, or multi-videos sequences, as produced by the Grimage platform at the INRIA Grenoble. The objective is to build the data sequences required to conduct research activities in the field of 4D modelling.

Teams :
Morpheo is a joint research team of the INRIA Grenoble Rhne-Alpes and of the Laboratoire Jean Kuntzmann, a joint research unit of CNRS, INRIA and Grenoble Universities. Morpheo is located at the INRIA Grenoble Rhne-Alpes in Montbonnot, France. Morpheos main objective is the ability to perceive and to interpret moving shapes using multiple camera systems for the analysis of animal motion, animation synthesis and immersive and interactive environments. Multiple camera systems allow dense information on both shapes and their motion to be recovered from visual cues. Such ability to perceive shapes in motion brings a rich domain for research investigations on how to model, understand and animate real dynamic shapes.

Dates :
starting date : 14 January, 2014
ending date : 21 January, 2014

Facilities descriptions :
http://visionair-browser.g-scop.grenoble-inp.fr/visionair/Browser/Catalogs/GRIMAGE.FR.html

Recordings & Results :
Tracking 3D human motion or estimating 3D human poses has long been studied in computer vision and graphic communities. The former assumes a human-like surface accessible and deform the surface according to 2D (silhouette) or 3D (visual hull) observations, whereas the latter starts with human features in a single image and infer 3D poses via learning approaches. Some good results have been reported in previous publications. However, most experiment datasets assume that background is properly subtracted. There are only objects of interest in the scene and sometimes even only one character. None of these methods discuss the performance when observations are noisy or contain outliers, nor is any public available dataset designed in this purpose.

Conclusions :
There are around 10 actors performing 5 different ac- tions: walk, sit, hammer, cut, reposition. Each actions is performed twice. The first time we put a chair as the irrelevant object, while the second time we put a table as a bigger occlusion. In a learning-based approach one can utilize the former as training set and test with the latter. In all there are around 90 sequences recorded; each consists of approximately 100 frames. Figure 1 shows a example of them. We notice that background is not properly subtracted in the first view and lead to a noisy silhouette image. Those artifacts however, does not appear in the visual hull thanks to robustness of voxel carving. Chair remains as desired in both silhouettes and visual hull because it is not learned as background.




Project Images :

img1.png



Other project resources :

fig1.pdf


.



Visionair logo

VISIONAIR / Grenoble INP / 46 avenue Felix Viallet / F-38 031 Grenoble cedex 1 / FRANCE
Project funded by the European Commission under grant agreement 262044