Plenoptic Face Reconstruction

 

 

Description

 

Plenoptic cameras [Ng05,Perwass12] are capable of imaging the scene from different perspectives. The information of the different perspectives is stored on a single image sensor which allows to acquire dynamic scenes and perform 3D reconstruction easily [Monteiro16]. The acquisition on a single sensor has the advantages described previously but it also limits the spacing between the viewpoints of the different perspectives which limits the field of view and the 3D reconstruction. Thus, in this work we plan to work with a stereo of plenoptic cameras.

 

3D face models are widely used for several purposes, such as biometric systems, face verification, facial expression recognition, 3D visualization, and so on. However, the reconstructed models of the face generated from the data of the optical setups used are quite noisy due to the lack of texture and the thin structures present in the face. In this work, we want to improve the reconstruction using a stereo of plenoptic cameras complemented with view interpolation [Vagharshakyan17]. Additionally, one wants to use symmetric priors and methods like the one presented in [Lourenço18] to improve the thin and textureless regions reconstruction.

 

Objectives

 

- Study plenoptic camera

- Acquire real dataset for face reconstruction

- Perform view interpolation (https://videos-rennes.inria.fr/video/r1S20C_Lr 27:42 a 28:06) and 3D face reconstruction (https://videos-rennes.inria.fr/video/HJ5KVkrUH 43:20 a 43:47)

 

References:

 

[Ng05] Ng, Ren. Digital light field photography. Stanford, CA: stanford university, 2006.

[Perwass12] Perwass, C., & Wietzke, L. (2012, February). Single lens 3D-camera with extended depth-of-field. In Human Vision and Electronic Imaging XVII (Vol. 8291, p. 829108). International Society for Optics and Photonics.

[Monteiro16] Monteiro, Nuno Barroso, Joao Pedro Barreto, and José Gaspar. "Dense lightfield disparity estimation using total variation regularization." International Conference on Image Analysis and Recognition. Springer, Cham, 2016.

[Lourenço18] Lourenco, Rui, et al. "Silhouette enhancement in light field disparity estimation using the structure tensor." 2018 25th IEEE International Conference on Image Processing (ICIP). IEEE, 2018.

[Vagharshakyan17] Vagharshakyan, Suren, Robert Bregovic, and Atanas Gotchev. "Light field reconstruction using Shearlet transform." IEEE transactions on pattern analysis and machine intelligence 40.1 (2017): 133-147.

 

 

More information:

 

Conventional cameras give information about the total amount of light that reaches each position of the sensor. During image formation, there is some information of the light field that is lost, for example: (i) the direction of the ray and; (ii) the contribution of each ray to the total amount of light captured in an image.

 

Plenoptic cameras like Lytro and Raytrix preserve this information. The additional information allows retrieving 3D information from the scene using a single image. Reconstruction methods for plenoptic images are based on identifying line features on epipolar plane images or on shearing the lightfield according to some customized metric for feature matching. These algorithms are computationally demanding and are still not capable of giving accurate depth estimations for low-textured and occluded regions.

 

On the other hand, depth sensors are active sensors that project a structured light pattern into the scene. This allows obtaining depth measurements regardless of the texture of the scene. Nonetheless, standard depth sensors give depth estimates with lower accuracy in the near range (less than 1.5 m).

 

References:

 

[Dansereau13] Dansereau, Donald G., Oscar Pizarro, and Stefan B. Williams. "Decoding, calibration and rectification for lenselet-based plenoptic cameras." Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on. IEEE, 2013.

 

[Herrera12] Herrera, Daniel, Juho Kannala, and Janne Heikkilä. "Joint depth and color camera calibration with distortion correction." IEEE Transactions on Pattern Analysis and Machine Intelligence 34.10 (2012): 2058-2064.

 

 

Place for conducting the work-proposal:

 

ISR / IST

 

 

 

More MSc dissertation proposals on Computer and Robot Vision in:

http://users.isr.tecnico.ulisboa.pt/~jag/msc/msc_2018_2019.html