To benefit the community, we release the dataset that we utilized for real-time event-based dense mapping. This dataset also serves as a basis for comparing our event-based dense mapping approach with data captured from a depth camera.
Our handheld device includes a power supply unit, an onboard computer NUC (equipped with Intel i7-1260P, 32GB RAM, and Ubuntu 20.04 operation system), a DAVIS346 event camera, and a Intel® RealSense™ D455 RGB-D camera.
We release the all of our schematic files with STL files format, which can be imported and printed directly. Moreover, we also release the CAD source files (with suffix “*.SLDPRT and *.SLDASM”), which can be opened and edited with Solidworks.
Please note that the depth maps provided by the depth camera only serve as references for qualitative comparison.
| Sequence Name | Collection Date | Total Size | Rosbag (Baidu Disk) |
|---|---|---|---|
| HKU_LG_office_1 | 2023-11 | 8.91 GB | Rosbag |
| HKU_LG_office_2 | 2023-11 | 7.71 GB | Rosbag |
| HKU_LG_office_3 | 2023-11 | 9.49 GB | Rosbag |
| HKU_LG_office_4 | 2023-11 | 8.55 GB | Rosbag |
| HKU_LG_office_5 | 2023-11 | 8.01 GB | Rosbag |
| HKU_LG_factory_1 | 2023-11 | 6.13 GB | Rosbag |
| HKU_LG_factory_2 | 2023-11 | 5.98 GB | Rosbag |
| HKU_LG_factory_3 | 2023-11 | 6.33 GB | Rosbag |
| HKU_Logo_wall_1 | 2023-11 | *** GB | Rosbag |
| HKU_Logo_wall_2 | 2023-11 | *** GB | Rosbag |
| HKU_Logo_wall_3 | 2023-11 | *** GB | Rosbag |
The video demos of evaluating EVI-SAM using this dataset are available on Bilibili 1 2 3 . The global mapping performance of our EVI-SAM is also illustrated in the following figur, showing the surface mesh generated through TSDF-based map fusion. Our global event-based dense mapping exhibits excellent global consistency. The supplemental video also illustrates the incremental reconstruction of the event-based surface mesh from updated TSDF voxels. This process enables on-demand meshing for visualization, allowing flexibility in generating the mesh at any time. Our TSDF-based map fusion for global mapping is designed to generate surface meshes that enable humans to assess the 3D reconstructed environment more effectively. This capability supports high-level mission goals, such as collision-free motion planning. During evaluation, an onboard computer NUC is utilized to support real-time pose estimation and local event-based dense mapping. However, the NUC lacks sufficient computational power to support a real-time meshing process. Therefore, we utilize a personal computer (Intel i7-11800H, 32GB RAM) without GPU to output the global mesh of EVI-SAM for the global mapping evaluation.
Visualization of the estimated camera trajectory and global 3D reconstruction (surface mesh) of our EVI-SAM. Sequentially display from right to left includes the event-based dense point clouds with texture information and intensity images, at selected viewpoints. (click the image to open video demo)
- We also use this dataset to evaluate some traditional image-based dense mapping methods using monocular, stereo, and RGB-D setup.
- A project of monocular dense mapping using RGB camera with VINS-MONO as pose estimation can be seen in Github Repository.
- We also use this dataset to evaluate some NeRF-based SLAM works, such as Nice-SLAM, Co-SLAM, NeRF-SLAM, and InStant-NGP+COLMAP, etc.
- We provide a toolbox that converts the rosbag into TUM-data format.