Table 4.
Performance comparison on the MVSEC dataset (indoor sequences), showing per-sequence and total average end-point error in pixels per second.
| Model | indoor_flying1 | indoor_flying2 | indoor_flying3 | AEE sum |
|---|---|---|---|---|
| EV-FlowNet (Zhu et al., 2018b) | 1.03 | 1.72 | 1.53 | 4.28 |
| Zhu et al. (2019) | 0.58 | 1.02 | 0.87 | 2.47 |
| Spike-FlowNet (Lee et al., 2020) | 0.84 | 1.28 | 1.11 | 3.23 |
| Back to Event BasicsEvf (Paredes-Vallés and de Croon, 2021) | 0.79 | 1.40 | 1.18 | 3.37 |
| Back to Event BasicsFire (Paredes-Vallés and de Croon, 2021) | 0.97 | 1.67 | 1.43 | 4.07 |
| XLIF-EV-FlowNet (Hagenaars et al., 2021) | 0.73 | 1.45 | 1.17 | 3.35 |
| XLIF-FireNet (Hagenaars et al., 2021) | 0.98 | 1.82 | 1.54 | 4.34 |
| Orchard et al. (2021) | 0.83 | 1.22 | 0.97 | 3.02 |
| Fusion-FlowNet (Lee et al., 2022) | 0.56 | 0.95 | 0.76 | 2.27 |
| Adaptive-SpikeNet (best ANN) (Kosta and Roy, 2022) | 0.84 | 1.59 | 1.36 | 3.79 |
| Adaptive-SpikeNet (best SNN) (Kosta and Roy, 2022) | 0.79 | 1.37 | 1.11 | 3.27 |
| FSFNFP (Apolinario et al., 2022) | 0.82 | 1.21 | 1.07 | 3.10 |
| FSFNHP−ADC (Apolinario et al., 2022) | 0.85 | 1.29 | 1.13 | 3.27 |
| Shiba et al. (2022) | 0.42 | 0.60 | 0.50 | 1.52 |
| Ours | 0.58 | 0.72 | 0.67 | 1.97 |
Best result in bold, runner-up underlined. Starting from a model pre-trained on DSEC, we show state-of-the-art performance without modifying our pipeline.