Datasets:
The consistency of EV, aperture, and ISO
Hello,
This is an impressive dataset. However, I have a few remaining inquiries regarding its composition and data acquisition strategy.
Specifically, if the metadata for a data sample indicates an ISO of 100, does this imply that both EV and ISO are held constant across the entire group, with only aperture and shutter speed varying? Alternatively, does the ISO 100 value refer strictly to the f/22 source image, while other images in the set might utilize different ISOs to maintain a constant shutter speed and EV?
I noticed in the paper that the data capture process is completed within 2 seconds, which suggests that long exposures are avoided. However, I am curious about how noise is managed in low-light environments, particularly for shots taken at f/22.
To summarize my concerns regarding the trade-offs:
If [EV, ISO] are consistent: As the aperture stops down (f-number increases), the exposure time must increase. At f/22, this could potentially lead to long exposures, causing pixel shift or motion misalignment.
If [EV, Shutter Speed] are consistent: As the aperture stops down, the ISO must increase to maintain exposure. Consequently, f/22 shots would likely suffer from significant noise.
Could you kindly clarify how your data construction pipeline addresses these potential issues?
Thank you very much for your time and assistance.
Best regards
Hio,
Thanks for your interest.
ISO remains constant within a scene capture (as Aperture is specified exposure time is varied for EV to remain constant as you point out above).
I think for low light/night scenes we used a maximum of ISO 1600, which still produced a clean image.
The "2 seconds" in our manuscript at some point included the half sentence "under good lighting conditions", which was probably cut at some point during the rewriting process to fit the template...
As we used the first approach, low light scenes can have quite high exposure times (10s max for f22 if I remember correctly). Hence they are either completely static (statues, rocks, clear sky without clouds etc. where wind is less of an issue) or under more controlled indoor conditions. We have not observed issues with sensor pixel shift under these conditions.
Btw Sensor IBIS was turned off in general for all scenes as we noticed that it could cause pixel misalignment between individual scene captures!).
We also considered the second approach, but did not implement this as it would implicitly introduce a denoising task into our data.
Best Regards!