🥺 E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization

[NeurIPS 2025]

Wenpu Li*  ·  Bangyan Liao*  ·  Yi Zhou  ·  Qi Xu  ·  Pian Wan  ·  Peidong Liu†

Paper | Project Page

This repository is an official PyTorch implementation of the paper "E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization". We reavel that implicit regularizations can enable the mutual promotion of self-supervised learning for optical flow and egomotion. Importantly, our method is capable of simultaneously learning the optical flow field and camera motion solely from event data.

📢 News & 🚧 TODO

• Launch the paper presentation video..
• Publish the v2 paper on arXiv.
• Add point tracking visualization results.
• Use uv to manage project environments and dependencies.
• Release the training code for DSEC dataset
• 2025.10.15 Publish the v1 paper on arXiv.
• 2025.10.13 Training code for MVSEC dataset has been released.
• 2025.10.12 Our project homepage is available online.
• 2025.09.18 Our paper was accepted by NeurIPS2025!! Congratulates to all collaborators!!

⚙️ Installation

Working in progress...

📊 Data Preparation

We conducted experiments on the MVSEC dataset and DSEC dataset. After downloading the datasets, for MVSEC, you only need to modify the data_path and gt_path in the config file to your paths. For DSEC, you only need to modify the data_path and timestamp_path in the config file to your paths.

The download link is as follows:

• MVSEC: [hdf5] [gt flow]
• DSEC: [hdf5] [timestamps]

🚀 Run

You can run E-MoFlow on the MVSEC dataset and DSEC dataset by:

python train_on_mvsec.py --gpu <gpu_idx> --config <config_file_path>
python train_on_dsec.py --gpu <gpu_idx> --config <config_file_path>

Additionally, you can modify the config file to conduct ablation studies or enable early stopping strategies to achieve a trade-off between speed and accuracy.

After running, the following results will be output:

outputs/
├── project/
│   ├── expname/
|   |   |-- origin_iwe                  # Original IWE
|   |   |-- pred_iwe                    # Warped IWE using predicted optical flow
|   |   |-- pred_flow                   # Predicted optical flow
|   |   |-- submission_pred_flow        # Predicted optical flow for DSEC evaluation (only for DSEC)
|   |   |-- gt_flow                     # Ground truth optical flow (only for MVSEC)
|   |   |-- motion                      # Estimated egomotion (only for MVSEC)
|   |   |-- metric.txt                  # Quantitative results (only for MVSEC)
|   |   |-- time_stats.txt              # Runing time statistics
|   |   |-- early_stopping_stats.txt    # Early stopping statistics

📖 Citation

If you find our work useful, please consider citing our paper:

@inproceedings{li2025emoflow,
      author = {Wenpu Li and Bangyan Liao and Yi Zhou and Qi Xu and Pian Wan and Peidong Liu},
      title = {E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization},
      booktitle = {Annual Conference on Neural Information Processing Systems (NeurIPS)},
      year = {2025}
  }