| Metric | Gen-3 | IMX990 | CNN | ELLI-NOVA | |--------|-------|--------|-----|-----------| | PSNR (dB) | 18.2 | 22.1 | 28.4 | | | SSIM | 0.42 | 0.58 | 0.81 | 0.94 | | Color accuracy (ΔE2000) | N/A | 32.4 | 18.1 | 5.3 | | Dynamic range (stops) | 6 | 12 | 14 | 21 | | Latency (ms) | <1 | 45 | 32 | 11 |
[1] Ellington, A., et al. "Variational autoencoders for Poisson inverse problems." NeurIPS 2023. elli nova nvg
No real-world dataset exists for multi-spectral, low-photon video. We therefore generated a synthetic dataset using physics-based rendering (Mitsuba 3 + custom SPAD noise model) over 10,000 scenes: | Metric | Gen-3 | IMX990 | CNN