Unsupervised Video Interpolation Using Cycle Consistency

Explore advanced techniques in unsupervised video interpolation using cycle consistency in this 30-minute lecture from the University of Central Florida. Delve into frame interpolation methods, including sparse optical flow and bi-directional flow-guided interpolation. Examine the limitations of supervised architectures like Super SloMo and discover how unsupervised approaches overcome these challenges. Learn about cycle consistency GANs, time domain consistency constraints, and pseudo-supervised loss functions. Analyze experimental setups, datasets, and evaluation methods used to test low-resolution training, domain gap testing, and fine-tuning for domain transfer. Gain insights from qualitative results and ablation studies on optimal weights, enhancing your understanding of cutting-edge video interpolation techniques.

Intro
Overview
Frame Interpolation Techniques
Sparse Optical Flow
Common Frame Interpolation Implementation
Bi-directional flow-guided Interpolation
Super SloMo: Base Supervised Architecture
Limitations on the state-of-the-art
Unsupervised Video Interpolation
Unsupervised Interpolation
Cycle Consistency GAN
Cycle Consistency Cost Function
Time Domain Consistency Constraint
Pseudo Supervised Loss
Primary Objective Function
Final Training Loss Function
Dataset and Metrics
Experiment Setup: Datasets Used
Experiment Setup: Training
Evaluation Methods
Experiment: Low Resolution Unsupervised Training
Experiment: Domain Gap Testing
Experiment: Fine Tune Domain Transfer
Experiment: Qualitative Results
Ablation: Optimal Weights
Conclusion