Annonce

Les commentaires sont clos.

Calibrate the Depth Image for Salient Object Detection

23 Novembre 2021


Catégorie : Stagiaire


Supervisor: Zongwei Wu, Pr. Cédric Demonceaux (Imvia, Dijon, France)

Term: 6 month (to start ASAP)

Compensation: 573.50e per month

Device: Local GPU 1080ti + Remote V100s

Key-Words: Deep Learning, Calibrate Depth, RGB-D fusion, RGB-D SOD

Contact : zongwei_wu@etu.u-bourgogne.fr; cedric.demonceaux@u-bourgogne.fr

Recent RGBD-based models for saliency detection have attracted research attention. The depth clues such as boundary clues, surface normal, shape attribute, etc., contribute to the identification of salient objects with complicated scenarios. However, in the current RGB-D SOD dataset, the depth map is or acquired by a depth sensor which may contain measurement error, or estimated from classical methods from 10 years ago. As a result, the depth maps are often exceedingly noisy at the object boundaries. What is more, the foreground object often differs only slightly from the surrounding background in the depth maps. These issues severely limit the potential performance of RGB-D models.

The recent development of monodepth estimation has shown great success. An off-the-shelf depth prediction model can sometimes show better performance compared to the raw depth. Inspired by this observation, we want to profit from the « good » estimated depth as supervision to improve the raw depth.

 

 

Objectives:

We provide an estimated depth dataset with good quality. The intern needs to first do the lecture on the state-of-the-art (SOTA) monodepth estimation field or depth completion field. The first objective is to propose a lightweight monodepth model or depth completion model to improve the depth quality. The second objective is to evaluate the performance of SOTA RGB-D SOD models with the newly estimated depth. We should be able to observe an improved performance compared to the plain version.

The next step is to improve the RGB-D fusion model for SOD. It can be a lightweight model but with equivalent performance. Or a complex model that outperforms the SOTA performance.

All the research work carried out by the intern will open the door to scientific publications for international conferences and/or impact factor journals.

 

Reference:

Ji, Wei, et al. "Calibrated RGB-D Salient Object Detection." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

Zhang, Wenbo, et al. "Depth Quality-Inspired Feature Manipulation for Efficient RGB-D Salient Object Detection." Proceedings of the 29th ACM International Conference on Multimedia. 2021.

Wu, Zongwei, et al. "Modality-Guided Subnetwork for Salient Object Detection."2021 ninth international conference on 3D vision (3DV). IEEE, 2021.