E2HQV

Official Implementation for "E2HQV: High-Quality Video Generation from Event Camera via Theory-Inspired Model-Aided Deep Learning" - AAAI 2024 arxiv aaai

[Important] Successor: PIE-Net — the next generation better version of E2HQV with probabilistic intensity-event modeling (PIEM), per-pixel uncertainty, and a pip-installable package (pip install event-pienet[realtime]). Pretrained PIE-Net and PIE-Net-Lite weights are included; real-time demo supported.

This repository remains the official E2HQV (AAAI 2024) code and benchmark artifacts for reproducibility.

E2HQV Generated Video Frames for Benchmarking

To benchmark with our method without processing your own data, you can find E2HQV-generated frames for evaluation on Google Drive. Below are the model's statistics on each dataset and scene (following the initial evaluation protocal of EVREAL):

[Overall]

Method	IJRR MSE↓	IJRR SSIM↑	IJRR LPIPS↓	MVSEC MSE↓	MVSEC SSIM↑	MVSEC LPIPS↓	HQF MSE↓	HQF SSIM↑	HQF LPIPS↓
E2VID	0.212	0.424	0.350	0.337	0.206	0.705	0.127	0.540	0.382
FireNet	0.131	0.502	0.320	0.292	0.261	0.700	0.094	0.533	0.441
E2VID+	0.070	0.560	0.236	0.132	0.345	0.514	0.036	0.643	0.252
FireNet+	0.063	0.555	0.290	0.218	0.297	0.570	0.040	0.614	0.314
SPADE-E2VID	0.091	0.517	0.337	0.138	0.342	0.589	0.077	0.521	0.502
SSL-E2VID	0.046	0.364	0.425	0.062	0.345	0.593	0.126	0.295	0.498
ET-Net	0.047	0.617	0.224	0.107	0.380	0.489	0.032	0.658	0.260
E2HQV (Ours)	0.028	0.682	0.196	0.032	0.421	0.460	0.019	0.671	0.261

[IJRR]

	boxes_6dof	calibration	dynamic_6dof	office_zigzag	poster_6dof	shapes_6dof	slider_depth
MSE↓	0.0354	0.0206	0.0278	0.0214	0.0345	0.0407	0.0129
SSIM↑	0.5638	0.6471	0.7185	0.6802	0.5552	0.8194	0.7879
LPIPS↓	0.2574	0.1639	0.1965	0.2239	0.1978	0.1712	0.1623

[MVSEC]

	indoor_flying1	indoor_flying2	indoor_flying3	outdoor_day1	outdoor_day2
MSE↓	0.0235	0.0194	0.0224	0.0518	0.0403
SSIM↑	0.4495	0.4249	0.4484	0.3343	0.4462
LPIPS↓	0.4381	0.4444	0.4262	0.5802	0.4086

[HQF]

	bike_bay_hdr	boxes	desk	desk_fast	desk_hand_only	desk_slow	engineering_posters	high_texture_plants	poster_pillar_1	poster_pillar_2	reflective_materials	slow_and_fast_desk	slow_hand	still_life
MSE↓	0.0306	0.0139	0.0146	0.0087	0.0135	0.0223	0.0207	0.0280	0.0108	0.0084	0.0147	0.0246	0.0304	0.0225
SSIM↑	0.5689	0.7571	0.7358	0.7781	0.7485	0.6867	0.6537	0.5559	0.6195	0.6543	0.6924	0.6737	0.5779	0.6878
LPIPS↓	0.3532	0.1850	0.1808	0.1771	0.2842	0.2711	0.2444	0.2166	0.2746	0.2651	0.2403	0.2531	0.3629	0.2087

Generate Video Frames with the Trained E2HQV

[Important] Looking for the latest model? Use PIE-Net instead — same research line, improved architecture, pip install event-pienet, and a real-time camera demo. This section documents the original E2HQV workflow.

Fix on 06/27/2024: app.py line 144 replace the p_states to current_states: return rf0, f01.detach(), last_gt, current_states, all_output

Note: Due to the size limitation on GitHub, the complete code along with the model weights is stored on Google Drive.

• On Google Drive, we provide minimal code to predict video frames using event-streams represented as voxel grids with 5 temporal bins. This representation was proposed by Alex et al. in their CVPR 2019 paper.

• An example sequence of voxel grids can be found in ./dataset/desk_fast_voxelgrid_5bins_examples. To generate the corresponding frames, simply run python3 app.py in the terminal.

• If you wish to use E2HQV with your own event data, place your event temporal bins in the form of a 5xHxW numpy array saved in .npy format (to ./dataset/desk_fast_voxelgrid_5bins_examples). Then, execute python3 app.py to process your data. In the Dataset Preparation section, we will provide detailed instructions and the necessary code to convert raw event data into voxel format.

Known Issue: The training process did not utilize optical flow, unlike other methods such as E2VID. As a result, the temporal consistency is suboptimal.

To Cite

@inproceedings{qu2024e2hqv,
  title={E2HQV: High-Quality Video Generation from Event Camera via Theory-Inspired Model-Aided Deep Learning},
  author={Qu, Qiang and Shen, Yiran and Chen, Xiaoming and Chung, Yuk Ying and Liu, Tongliang},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={5},
  pages={4632--4640},
  year={2024}
}

Dataset Preparation

You can find the e2voxel_grid.py script for converting events to voxel grids in Google Drive. More python implementation for other common event representations (e.g., two-channel, four-channel, TORE, EvRep, and EvRepSL) can be found here.

To Be Updated