Stereo Omnidirectional Depth Estimation Dataset with LiDAR Labels

HELVIPAD provides 39,553 paired equirectangular images for omnidirectional stereo depth estimation. The dataset was created by author chcorbi and last updated in September 2025. It includes corresponding pixel-wise depth and disparity labels derived from LiDAR point clouds across diverse scenes.

Use Cases

• Train models to predict pixel-wise depth labels from paired equirectangular stereo images.
• Benchmark omnidirectional stereo matching algorithms using provided ground-truth disparity labels.
• Evaluate depth estimation model robustness across varying lighting conditions, including night-time scenes.
• Develop multi-task models that jointly learn from LiDAR-derived depth and image data for scene understanding.

Strengths

• Contains 39,553 paired equirectangular images, providing a substantial collection for training.
• Includes pixel-wise depth and disparity labels derived from LiDAR point clouds, offering precise ground truth.
• Spans diverse indoor and outdoor scenes under varying lighting conditions, including night-time environments.

Limitations

• The specific geographic origin of the captured scenes is not detailed, which may limit understanding of environmental bias.
• The dataset size of 39,553 image pairs may be moderate compared to larger-scale 2D image datasets for depth estimation.

Provenance

Source

huggingface.co/datasets/chcorbi/helvipad

Collection Method

Images captured using a top-bottom 360° camera setup; labels derived from LiDAR point clouds.

Freshness

Last updated on 2025-09-24.