Exploiting Illumination Decline: Advances in Monocular Visual SLAM, Self-Supervised Depth Estimation, and Watertight 3D Reconstruction

Abstract:

Despite today’s wide variety of sensors, some applications still rely on monocular imaging, limiting the ability to perform true-scale 3D reconstructions. This talk presents three pioneering steps towards real scale perception in Monocular Visual SLAM inside the human body. The first approach introduces a photometric model and sets the basis for exploiting colocalized lighting with a calibrated monocular endoscope. On a second step, we build on top of illumination decline mode to achieve single-view self-supervision for accurate depth estimation. This method achieves performance levels akin to supervised techniques, even in the absence of depth ground-truth data. The third stage unveils watertight 3D reconstruction of endoluminal cavities from sequences of images, offering exceptional accuracy and the potential for automatic quality assessment in cancer screening. These results herald how exploiting the illumination decline is key to unlocking the full potential of monocular VSLAM.

Short Bio:

Victor M. Batlle is completing his second year as a Ph.D. student at the University of Zaragoza in Spain under the supervision of Prof. Juan D. Tardos. This summer he is an intern at Apple’s ACVML Team in Santa Clara Valley. His research focuses on monocular 3D reconstruction, and aims to combine computer vision techniques with light transport principles inspired by computer graphics and simulation of appearance. His recent work has been accepted at IROS, ICCV, and MICCAI conferences.