sunmengxi/panorama_video_process_360x

This dataset is built from the 360-degree panorama videos in quchenyuan/360x_dataset_HR. We processed 231 of the 232 source videos; one video was excluded because the downloaded file was corrupted.

For each valid source panorama video, we generate a full set of perspective-view probe videos by crossing four factors:

• 3 durations: 30s, 60s, and 120s
• 3 viewing patterns:
  • target-away-target
  • away-target-away-target, with the same away view used before and after the target
  • away-target-away-target, with different away views before and after the target
• 2 difficulty levels, defined by the randomly sampled target-away ratio:
  • Easy: [0.2, 0.4]
  • Hard: [0.6, 0.8]
• 3 repeats per condition, distinguished by repeat_index

Thus, each valid source video produces 3 durations x 3 viewing patterns x 2 difficulties x 3 repeats = 54 processed videos.

Each output is rendered as a perspective view from the original panorama at 854x480 resolution and 24 FPS. The target view is randomly sampled by yaw angle, while the away view is randomly sampled about 118-122 degrees away from the target. Roll and Pitch kept to be zero. The horizontal field of view(fov) is fixed per output and randomly sampled from 95-105 degrees. Transitions between views are smooth with an average camera rotation speed at 60 deg/s.

The target-away ratio is computed only over the counted target/away interval, where transition frames and the final 5s prediction window are excluded from this ratio. Each sample ends with this prediction window, which includes the final rotation back to the target view and a short final target hold.