A photorealistic synthetic stereo dataset with ground-truth depth and 6-DoF poses

This data article presents a synthetic stereo vision dataset composed of synchronized RGB stereo images, ground-truth depth maps, and precise 6-DoF ground-truth poses generated using Unreal Engine 4 integrated with the AirSim simulation framework. Four indoor virtual scenes resembling industrial and enterprise-like environments were created, including structural elements such as pipes, pillars, walls, furniture, and occlusions. Multiple acquisition configurations were executed as closed-loop drone trajectories, resulting in a total of 36,838 stereo image pairs. The dataset includes multiple trajectory smoothness conditions, stereo baselines, and camera convergence configurations. All RGB images have a fixed resolution of 640 × 480 pixels and are provided alongside pixel-aligned depth maps and time-stamped ground-truth poses. The dataset also includes association files linking stereo images, depth maps, and poses, following formats commonly used in visual SLAM benchmarks. The acquisition is scripted and free of stochastic components, enabling benchmarking and development of stereo depth estimation, visual odometry, and SLAM algorithms under controlled and repeatable indoor conditions.