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..
V2V: Scaling Event-Based Vision through Efficient Video-to-Voxel Simulation
Authors: Hanyue Lou, Jinxiu Liang, Minggui Teng, Yi Wang, Boxin Shi
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Leveraging this efficiency, we train several video reconstruction and optical flow estimation model architectures on 10,000 diverse videos totaling 52 hours an order of magnitude larger than existing event datasets, yielding substantial improvements. |
| Researcher Affiliation | Academia | Hanyue Lou1,2 Jinxiu Liang3 Minggui Teng1 Yi Wang2,4 Boxin Shi1# 1 Peking University 2 Shanghai Innovation Institute 3 National Institute of Informatics 4 Shanghai AI Laboratory EMAIL EMAIL EMAIL |
| Pseudocode | No | The paper describes the V2V conversion process in Section 3.3 but does not present it as a formal pseudocode or algorithm block. |
| Open Source Code | Yes | $ Code available: https://github.com/HYLZ-2019/V2V |
| Open Datasets | Yes | We use the Web Vid [2] video dataset for model training... We prepare the ESIM-280 dataset... We performed zero-shot model evaluation on the real event datasets HQF [33] and EVAID [6]... More evaluation details (Section A), test results on IJRR [25] and MVSEC [43] (Section E)... |
| Dataset Splits | Yes | From the 2.5M videos of Web Vid, we randomly sample 10K, 1K and 100 videos, forming our datasets Web Vid10K, Web Vid1K and Web Vid100... We evaluate the methods on selected sequence cuts of IJRR [25], MVSEC [43], HQF [33], and EVAID [6]... we use the full HQF sequences, and cut the IJRR and MVSEC sequences with boundaries as listed in Table 4. The same MVSEC sequence cuts are used for optical flow evaluation. The frames provided in the EVAID [6] dataset have relatively better visual quality. To reduce the testing burden, we only crop a 5-second segment (up to 750 frames, as EVAID has a high frame rate) from each of the longer sequences. The boundaries of these segments are listed in Table 5. |
| Hardware Specification | No | All our experiments were conducted on single-GPU instances, and the peak GPU memory usage does not exceed 80 GB. |
| Software Dependencies | No | The paper mentions using skimage.metrics.structural_similarity for SSIM calculation and RAFT models, but does not provide specific version numbers for any software libraries or frameworks used in the implementation. |
| Experiment Setup | Yes | We train the E2VID model with batch size 12, crop size 128 128 and constant learning rate 0.0001. For ETNet and Hyper E2VID, we also follow their original training protocols. The training epochs ( Eps in Table 2) are set to keep the total amount of iterations approximately the same across experiments. For the event threshold parameters c+ and c used in the V2V simulator, we first uniformly sample a threshold c from the range [0.05, 2]... We combine an L1 loss with the VGG LPIPS loss... In Table 2, we refer to the loss combination of VGG + L1 + Temporal-Consistency-Half as V+L1+H. |