25 sequences · 255GB · 6.9h · 2.8km
Passive, active, and neuromorphic vision with IMU, DVL, and pressure sensors for robust perception
High-fidelity dense point clouds
acquired by a 3D laser scanner
for mapping evaluation
Motion capture system (indoor) and INS-SfM fusion (outdoor) for localization evaluation
Model: FindROV H20
Weight: 23 kg
Max Depth: 300 m
Velocity: 2 knot
Cable: 100 m
Model: 3DM-CV5-25
Freq: 100 Hz
R/P: ±0.5°
Yaw: ±1°
Mag: ±2.5 Gauss
Model: A50
Beam: 4-beam
Angle: 22.5°
Freq: 4–26 Hz
Res: 0.1 mm/s
Model: B30
Range: 0–30 bar
Depth: 300 m
Acc: ±200 mbar
Res: 0.2 mbar
View: Front
USB RGB
Res: 1920×1080
Freq: 30 Hz
FoV: 80°×64°
View: Down
Model: D435i
Res: 640×480
Freq: 30 Hz
FoV: 87°×58°
Model: DAVIS346
Resolution: 346 × 260
Temporal: 1 μs
Latency: < 1 ms
Dynamic Range: 120 dB
Freq: 70 Hz
Points: 512
λ: 450 nm
Power: 3 W
Angle: 35°
Scene-wise characteristics in the AquaTriV dataset.
We provide multi-modal raw sensory data including active vision, passive vision, and neuromorphic vision streams, along with synchronized navigation measurements. Below shows representative visual samples and ROS topic formats.
Examples of visual raw images in the AquaTriV dataset.
| Type | Topic Name | Message Type | Rate (Hz) |
|---|---|---|---|
| Active Vision | /scanner/left_image/compressed | sensor_msgs/CompressedImage | 100 |
| /scanner/right_image/compressed | sensor_msgs/CompressedImage | 100 | |
| Passive Vision | /stereo/infra1/compressed | sensor_msgs/CompressedImage | 30 |
| /stereo/infra2/compressed | sensor_msgs/CompressedImage | 30 | |
| /stereo/monocular/compressed | sensor_msgs/CompressedImage | 30 | |
| /stereo/imu | sensor_msgs/IMU | 200 | |
| Neuromorphic Vision | /dvs/events | dvs_msgs/EventArray | 30 |
| /dvs/image_raw/compressed | sensor_msgs/CompressedImage | 30 | |
| /dvs/event_rendering/compressed | sensor_msgs/CompressedImage | 30 | |
| /dvs/imu | sensor_msgs/IMU | 200 | |
| AHRS | /imu/data | sensor_msgs/Imu | 200 |
| /imu/mag | sensor_msgs/MagneticField | 50 | |
| DVL | /dvl/data | dvl_a50_2/DVL | 12 |
| /dvl/velocity | sensor_msgs/Imu | 12 | |
| /dvl/position | sensor_msgs/Imu | 4 | |
| Pressure | /pressure_sensor/depth | sensor_msgs/FluidPressure | 60 |
| Bonus | /robotview/monocular | sensor_msgs/CompressedImage | 30 |
| /waterview/monocular | sensor_msgs/CompressedImage | 30 |
We evaluate classical and multi-modal visual-inertial SLAM systems including VINS-Fusion, SVIn2, VD-Fusion, AEVINS and AquaSLAM. These methods represent state-of-the-art pipelines combining stereo vision, IMU, DVL, and event cameras.
Five baseline method localization evaluation results in the AquaTriV dataset.
Metrics include APE, ARE, RPE, RRE for accuracy, and CRS/TCR for robustness and tracking continuity.
We compare representative dense mapping and neural reconstruction baselines, including DR-Scan, DR-Sweep, DROID-SLAM, and MASt3R-SLAM, covering both geometric and learning-based approaches.
Four baseline method dense mapping evaluation results in the AquaTriV dataset.
Evaluation metrics include nearest neighbor error (NNE), completeness (COM), accuracy (AC), Chamfer distance (CD), and mesh metric error (MME).
Recent neural rendering approaches, particularly 3D Gaussian Splatting (3DGS), require stable and sufficiently dense geometric priors, which are often lacking in underwater datasets relying on sparse SfM-based reconstructions. By deploying SeaSplat on AquaTriV and leveraging dense laser-scanned point clouds for initialization, we achieve stable optimization and high-fidelity photo-realistic reconstruction across diverse underwater scenes.
