Since our checkpoint cannot pass Google Drive's virus check, we no longer provide scripts for automatic downloading from Google Drive (which would result in a Google warning instead of the checkpoint).
For details, please refer to issue#18.
However, you could not only download the checkpoint from Google Drive but our Github Release as well.
The data naming convention consists of the following components: {Method}_{Phase}_{Dataset}_{Camera Model}_{Training Setting}_Setting_Ratio{Range}.
For the LED method, there are two phases in total: Pretrain and Deploy. For Pretrain, the checkpoint obtained cannot be directly tested on any dataset; it is only used for subsequent fine-tuning. On the other hand, for all methods, the Deploy means that the checkpoint contains parameters consistent with the UNet used in the SID method, making it directly suitable for testing.
Regarding the training setting, there are two mainstream settings known as ELD (CVPR20) and PMN (MM22) settings. We represent these two settings as "CVPR20" and "MM22," respectively.
e.g. LED_Deploy_SID_SonyA7S2_CVPR20_Setting_Ratio100-300:
- "LED_Deploy": This refers to the LED method in the "deploy" phase.
- "SID_SonyA7S2": This indicates that the testing is done on the SonyA7S2 subset of the SID dataset.
- "CVPR20_Setting": This means that the training strategy during the "pretrain" phase is the same as the one used in the "ELD (CVPR20)" setting.
- "Ratio100-300": This indicates the range of the ratio is from 100 to 300.
The data naming convention consists of the following components: {Type}_{Noise Model}_{Noise Type}_{Camera Model}.
e.g. VirtualNoisyPairGenerator_ELD_ptrqc_5VirtualCameras.pth denotes the VirtualNoisyPairGenerator with ELD noise model and shot (poission), read (tukey lambda), row, quant noise and color bias (black level error). Also the checkpoint contains 5 random sampled virtual cameras.