Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
EPA: Boosting Event-based Video Frame Interpolation with Perceptually Aligned Learning
Authors: Yuhan Liu, LingHui Fu, Zhen Yang, Hao Chen, Youfu Li, Yongjian Deng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that this approach yields interpolated frames more consistent with human perceptual preferences. |
| Researcher Affiliation | Academia | 1Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University, 361005, P.R. China 2College of Computer Science, Beijing University of Technology 3Key Lab of Computer Network and Information Integration, Southeast University 4Department of Mechanical Engineering, City University of Hong Kong, Kowloon, Hong Kong SAR 5City U Shenzhen Research Institute, Shenzhen, P.R. China |
| Pseudocode | No | The paper describes the proposed EPA framework in detail with pipeline diagrams (Figure 2 and Figure 3) and textual descriptions of modules and processes. However, it does not include a formally labeled 'Pseudocode' or 'Algorithm' block with structured steps. |
| Open Source Code | Yes | Codes are available at https://github.com/yuhan0802/EPA. |
| Open Datasets | Yes | The synthetic benchmarks include Vimeo90k-Triplet [47] and GOPRO [33]. For real-world evaluation, we use HS-ERGB [40], comprising 15 scenes and various motion types; BS-ERGB [41], characterized by noise and complex non-rigid deformations; and Event Aid-F [7], containing various motion scenarios. |
| Dataset Splits | Yes | The entire model is trained on GOPRO [33] following [31], where synthetic event data is generated using the v2e simulator [9]. For data augmentation, both input frames and their corresponding event voxel grids are cropped to 256 256 and randomly augmented with rotation and flipping. |
| Hardware Specification | No | The paper's |
| Software Dependencies | No | Our method is optimized using Adam W [30] for 100 epochs within the Py Torch [36]. |
| Experiment Setup | Yes | For the proposed EPA, we first optimize the style adapter and reconstruction generator modules, after which their weights are frozen to train the bidirectional feature alginement module. In the first training stage, our method is optimized using Adam W [30] for 100 epochs within the Py Torch [36]. The initial learning rate is set to 1 10 4 and is gradually decreased to 1 10 6 via cosine annealing. The batch size is set to 40 for each training step. In the second stage, the bidirectional adaptation module is trained for 40 epochs under the same configuration. The entire model is trained on GOPRO [33] following [31], where synthetic event data is generated using the v2e simulator [9]. For data augmentation, both input frames and their corresponding event voxel grids are cropped to 256 256 and randomly augmented with rotation and flipping. Our normalized flow generation module follows the setup used in [45]. ... where Ξ»lap and Ξ»2 are set as 0.2, 0.1 respectively. |