confusion with the denoising process in TRT-NLOS

[ExpertsInEatting]

Are the original bike dataset measurements already noisy? For the light denoiser, what are considered the ground truth and the input? My understanding is that the noise is manually added during the model’s forward pass, and the ground truth is simply the raw transient data from the original bike dataset. Is this correct?
https://github.com/Depth2World/NLOST/blob/019c229432d05bdd6d1224a10f61254280451250/models/nlost.py#L93

[Depth2World]

Apologies for the delayed response — I missed your message earlier.

The transient measurements are first degraded using the noise generation tool provided in [1], and the noisy measurements are then used as the network input. In other words, the noise is not added by a hand-crafted function during the model forward pass.

So the ground truth is the original clean transient data from the bike dataset, and the input is the corresponding noisy transient measurement generated by the tool in [1].

[1] A computational model of a single-photon avalanche diode sensor for transient imaging

[batman47steam]

Does the '_spad' suffix in the bike dataset mean it's the noisy version, and the one without the suffix is the clean version?

[Depth2World]

yeap