Related Experiment Video
Updated: Dec 23, 2025
10:16
Digital Inline Holographic Microscopy DIHM of Weakly-scattering Subjects
Published on: February 8, 2014
12.6K
Deep Slow Motion Video Reconstruction With Hybrid Imaging System
Summary
This study introduces a novel deep learning system to create high-resolution slow motion videos from standard and high-frame-rate inputs. The method enhances video frame interpolation for realistic motion, overcoming limitations of current techniques.
Area of Science:
- Computer Vision
- Deep Learning
- Video Processing
Background:
- High-resolution slow motion video requires specialized high-speed cameras.
- Current frame interpolation methods assume linear motion, failing in complex scenarios.
Purpose of the Study:
- To reconstruct high-resolution slow motion video from hybrid inputs (standard and high-frame-rate).
- To improve frame interpolation accuracy for non-linear object motion.
Main Methods:
- A two-stage deep learning system: alignment and appearance estimation.
- Utilizing auxiliary high-frame-rate video for temporal information and alignment.
- Employing a context and occlusion-aware network for appearance estimation.
Main Results:
- Successfully reconstructed high-resolution slow motion videos from hybrid inputs.
- Demonstrated high-quality results on diverse test scenes.
- Showcased practical performance on real dual-camera setups.
Conclusions:
- The proposed system effectively generates high-resolution slow motion video.
- This approach overcomes limitations of linear motion assumptions in frame interpolation.
- The method offers a practical solution for enhancing video frame rates.
