Kernekoncepter
This paper proposes a comprehensive benchmark for evaluating video frame interpolation methods. The benchmark includes a carefully designed synthetic test dataset that adheres to the constraint of linear motion, consistent error metrics, and an in-depth analysis of the interpolation quality with respect to various per-pixel attributes such as motion magnitude and occlusion.
Resumé
The paper presents a benchmarking framework for evaluating video frame interpolation methods. The key aspects are:
Directory:
Introduction
Video frame interpolation is an increasingly popular research area with various applications
Existing test datasets and evaluation metrics are inconsistent, making fair comparisons challenging
The paper proposes a dedicated benchmarking framework to address these limitations
Related Work
Overview of existing test datasets for frame interpolation
Limitations of these datasets, including violation of linearity constraint and lack of in-depth analysis
Dataset Generation
The benchmark uses synthetic data generated by composing real-world sprites and backgrounds
The synthetic data adheres to the constraint of linear motion
Analysis of motion magnitude and angle distribution compared to existing datasets
Evaluation Metrics
Discussion of the PSNR metric and the proposed PSNR* definition to address flaws in the standard PSNR computation
Leveraging the synthetic data to analyze interpolation quality with respect to motion magnitude, angle, and occlusion
Proposal of a new PSNR*σ metric to evaluate temporal consistency in multi-frame interpolation
Submission Page
Description of the submission website that ensures consistent and comparable results
Features like computational efficiency evaluation and anomaly detection
Results
Quantitative evaluation of 21 representative frame interpolation methods across multiple resolutions
Analysis of the interpolation quality with respect to motion magnitude, angle, and occlusion
Multi-frame interpolation evaluation and computational efficiency assessment
Limitations
Focus on two-frame input interpolation, omitting other areas like non-linear or event-based interpolation
Inability to use patch-wise metrics due to the per-pixel analysis
Potential impact of ensembling on the benchmark results
Conclusion
The proposed benchmark is expected to benefit the frame interpolation community by providing new insights and accelerating research progress.