DCEGen: Dense Clutter Environment Generation Tool for Autonomous 3D Exploration and Coverage Algorithms Testing

How good MVSNets are at depth fusion

“…The depth maps allow for higher level of detail compared to voxel grids and possess natural connectivity compared to point clouds. Depth map representation requires an additional lossy fusion procedure to produce the final reconstruction, however the current state of the art in MVS is mostly held by the methods using this representation 1 .…”

“…To resolve this issue, the authors of R-MVSNet [8] use convolutional gated recurrent unit for cost volume regularization instead of 3D CNN, which allows them to reduce the growth of memory consumption from cubic to quadratic. In parallel, 1 according to tanksandtemples.org/leaderboard/AdvancedF arXiv:2011.14761v1 [cs.CV] 30 Nov 2020 the authors of Point-MVSNet [9] propose a coarse-to-fine strategy to achieve computational efficiency. They predict a lowresolution 3D cost volume to obtain a coarse depth map and iteratively upsample and refine it.…”

“…State-of-the-art MVS methods produce accurate 3D reconstructions from a set of photos captured from different points of view, but require distinctive texture for robust matching, and accurate respective camera parameters, that include positions, orientations and distortions. The methods can be applied within the mobile robot based 3D reconstruction field [1].…”

Robot-Assisted Language Learning: Scientific Data Analysis

“…Robotics is closely related to artificial intelligence and also applies the results of progress in this area [2]. Thanks to artificial intelligence, it is possible to generate a three-dimensional space for robot testing in simulation [3], generate a maze [4] and create obstacles to further calibrate algorithms for complex tasks [5]. Robots, thanks to machine learning algorithms, are able to navigate in the Gazebo simulation environment [6], [7], in real-world environments [8], and can also follow humans [9] and independently plan a route to move [10].…”