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..
E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization
Authors: Wenpu Li, Bangyan Liao, Yi Zhou, QiXu, Pian Wan, Peidong Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct comprehensive evaluations on the MVSEC dataset [64], which is the de facto standard dataset used in prior work to benchmark optical flow and 6-Do F ego-motion estimations. This dataset contains both indoor sequences recorded by drones and outdoor sequences recorded by vehicles. Additionally, we benchmarked optical flow estimation on the more challenging DSEC [65] dataset... For optical flow evaluation, we compute three standard metrics: endpoint error (EPE), angular error (AE) and the percentage of pixels with EPE > 3 pixels... |
| Researcher Affiliation | Academia | Wenpu Li1 , Bangyan Liao1,2 , Yi Zhou3, Qi Xu1,4, Pian Wan5, Peidong Liu1 , 1Westlake University 2Zhejiang University 3Hunan University 4Wuhan University 5Georgia Institute of Technology |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are present in the paper. The methodology is described using mathematical equations and textual descriptions. |
| Open Source Code | No | In the experiments section, we provide a detailed description of our implementation, and we will release the source code upon acceptance. (from NeurIPS Paper Checklist - 4. Experimental result reproducibility) We will release the code upon acceptance. (from NeurIPS Paper Checklist - 5. Open access to data and code) |
| Open Datasets | Yes | We conduct comprehensive evaluations on the MVSEC dataset [64], which is the de facto standard dataset used in prior work to benchmark optical flow and 6-Do F ego-motion estimations... Additionally, we benchmarked optical flow estimation on the more challenging DSEC [65] dataset... |
| Dataset Splits | No | Following prior works [17] [43], we partition the entire event sequence into multiple segments during training, with each segment containing 30k events for MVSEC [64] and 300k events for DSEC [65]. The paper does not explicitly specify the training, validation, and testing splits used for the MVSEC and DSEC datasets. |
| Hardware Specification | Yes | All experiments were conducted on a NVIDIA RTX 4090. |
| Software Dependencies | No | The paper mentions using AdamW optimizers and an Euler solver but does not provide specific version numbers for software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | The neural implicit flow field adopts the MLP architecture... The cubic spline modeling continuous camera motion employs only 4 control knots, whose 6-dimensional vectors are initialized to a constant value of 0.2... The weight of the differential geometric loss is set to 0.25 and the differential flow loss to 1. We employ two separate Adam W optimizers [67]... For the MVSEC dataset, the learning rate for the flow field is exponentially decayed from 1 10 4 to 6.3 10 5, whereas for the DSEC dataset, it is cosine annealed from 2 10 3 to 1 10 7. The learning rate for the camera motion is kept constant at 1 10 3 for both datasets. Each short event segment is trained for 1k iterations. |