We have developed a comprehensive dataset for object detection and 6D pose estimation in both underwater and dry environments, comprising both synthetic and real-world data. The dataset focuses on four objects of interest for underwater intervention.

An example showing the environmental conditions of intervention.

• RGB
• Depth Map
• Mask for single object
• CAD object models
• 2D ground truth labels
• 6D ground truth labels (rotation matrix 3x3 and translation vector 1x3)

• 33920 sysntetic images
• 87100 real-world images
• Resolution 640 x 480

• 4 objects
• Syntetic data:
• different sandy background textures
• color lights change randomly, with values chosen in specific ranges to avoid black lights
• lights directions change randomly (lights change their directions starting from the camera position for max 80°)
• object poses are chosen randomly
• Real data:
• 10 different scenes
• 42.5% of images acquired in 4 different underwater conditions
• 57.5% images acquired in 6 different dry scenes

• 1: Box
• 2: Cup
• 3: Jug
• 4: Hotstab