63,480 video triplets categorized into three distinct video processing tasks: frame interpolation, super-resolution, and denoising. The dataset provides multi-frame sequences designed to train neural networks that learn motion estimation specifically for image reconstruction and enhancement.
Use Cases
- Train video frame interpolation models using the 3-frame triplet sequences to synthesize intermediate frames
- Develop video super-resolution algorithms by mapping low-resolution input frames to high-resolution ground truth targets
- Implement video denoising pipelines using the noisy frame inputs and clean ground truth references
- Evaluate task-specific motion estimation by comparing task-oriented flow outputs against standard optical flow benchmarks
Strengths
- 63,480 video triplets provided for the frame interpolation sub-task
- Includes dedicated subsets for 4x super-resolution and video denoising with high-quality ground truth
- Data structured as 7-frame sequences or 3-frame triplets to support temporal context modeling
- Ground truth frames included for every sequence to facilitate end-to-end training of enhancement models