Table 5.
Performance comparison on the MVSEC dataset (outdoor sequences), showing average end-point error in pixels per second.
| Model | outdoor_day1 (px/s) |
|---|---|
| EV-FlowNet (Zhu et al., 2018b) | 0.49 |
| Zhu et al. (2019) | 0.32 |
| ECNmasked (Ye et al., 2020) | 0.30 |
| Spike-FlowNet (Lee et al., 2020) | 0.49 |
| Back to Event BasicsEvf (Paredes-Vallés and de Croon, 2021) | 0.92 |
| Back to Event BasicsFire (Paredes-Vallés and de Croon, 2021) | 1.06 |
| XLIF-EV-FlowNet (Hagenaars et al., 2021) | 0.45 |
| XLIF-FireNet (Hagenaars et al., 2021) | 0.54 |
| Fusion-FlowNet (Lee et al., 2022) | 0.59 |
| Adaptive-SpikeNet (best ANN) (Kosta and Roy, 2022) | 0.48 |
| Adaptive-SpikeNet (best SNN) (Kosta and Roy, 2022) | 0.44 |
| FSFNFP (Apolinario et al., 2022) | 0.51 |
| FSFNHP−ADC (Apolinario et al., 2022) | 0.48 |
| Shiba et al. (2022) | 0.30 |
| Ours | 0.85 |
Best result in bold, runner-up underlined. While far from the top performing contributions, our base pipeline is able to learn to estimate optical flow from scratch, without any optimization to make it tailored to the dataset and camera.