At MDS group, we focus on
Artificial Intelligence and Perception for the Marine Sciences and Environments, in particular at
the interface of Computer Vision, Computer Graphics, Robotics and Machine Learning. We therefore
need to deal with challenging but very interesting real-world images acquired by deep-sea robots.
Highlights are the major works.
Underwater cameras suffer from geometric distortion induced by refraction of light rays at the
water-glass-air interfaces. There has been works going on to analyze and calibrate such systems.
However, we ask the question: what are you going to do with the calibrated refractive cameras? Can
you use it in structure-from-motion? We therefore integrate the refractive geomtry within the COLMAP
framework and release it as open-source.
A semi-hierarchical approach which combines the good parts of incremental and global
structure-from-motion for large-scale AUV-based monocular camera seafloor reconstruction. Weak-area
revisiting helps to mitigate
inconsistencies in the reconstructed camera poses and map.
How to establish 3D models of partially translucent organisms? We adapt Differentiable Raytracing
to real data to: 1. capture Tomography-like internal structure; 2. establish relightable models; 3.
immerse models in other media; 4. exhibit high novel-view stability
ISPRS Journal of Photogrammetry and Remote Sensing
The BubbleBox project aims to develop a high-speed stereo camera system that can be deployed on the
seafloor of the deep ocean to measure and quantify the amount of gas released from the seafloor into
the water column. Checkout the livestream of
the BubbleBox deployment!
We discover that the camera model of a focus stacked image is essentially an affine model. We then
propose a macro-lens calibration approach based on this insight for 3D reconstruction of very tiny
objects using such cameras. This technology is used in the TechOceanS project.
3D reconstruction using a camera with an active light source creates an uneven and color-distorted
texture. We use physically-based differentiable ray-tracing to firstly estimate the light source in
a calibration scenario, and then optimize a medium- and light- free texture in the subsequent SfM
step.
In this work, we develop an effective method to align a camera with an underwater spherical port
(dome-port) for underwater camera systems. This technique is essential for the accurate calibration
of our underwater cameras.