search

ubicomplab / rPPG-Toolbox

rPPG-Toolbox: Deep Remote PPG Toolbox (NeurIPS 2023)
https://arxiv.org/abs/2210.00716
Other
459 stars 112 forks source link

Correct SNR calculation to reflect reference #288

Closed yahskapar closed 3 months ago

yahskapar commented 3 months ago

This PR makes some corrections to the SNR calculation in order to properly reflect the approach used in the CHROM paper, and as mentioned in the bottom of page six of this paper.

Using the PURE_UBFC-rPPG_TSCAN_BASIC.yaml for testing, here's results (including SNR) prior to corrections:

FFT MAE (FFT Label): 1.2974330357142858 +/- 0.3951013616658031 FFT RMSE (FFT Label): 2.8704957923240366 +/- 3.0489809798112115 FFT MAPE (FFT Label): 1.500155568072648 +/- 0.4672290826823667 FFT Pearson (FFT Label): 0.9890524988652464 +/- 0.02333192795742126 FFT SNR (FFT Label): 1.4945164762235956 +/- 1.1317032218020173 (dB)

And results after corrections:

FFT MAE (FFT Label): 1.2974330357142858 +/- 0.3951013616658031 FFT RMSE (FFT Label): 2.8704957923240366 +/- 3.0489809798112115 FFT MAPE (FFT Label): 1.500155568072648 +/- 0.4672290826823667 FFT Pearson (FFT Label): 0.9890524988652464 +/- 0.02333192795742126 FFT SNR (FFT Label): 6.641024545844431 +/- 1.1291389587742173 (dB)

This does seem to have a significant effect on the SNR, so I propose in a future iteration of the arXiv that all supplementary results that include SNR should be updated.