Training Configuration File for coco-rgb-aug Weights

[wangerniuniu]

Could you please point me to the correct training configuration file for thecoco-rgb-augmodel? If it's available in a repository, a link would be greatly appreciated.

[AliYoussef97]

Hello @wangerniuniu

Based on the FAQ here, coco-rgb-aug and pvgg-4 are the same model but different photometric augmentation during training.

I believe the default training in this repository leads to the coco-rgb-aug model, however the authors can correct me of course.

If we follow the configration tree, we start at the train-silk.yaml file, where the 3rd line has /models@model: silk-vgg as the default. If we go to the silk-vgg.yaml we can see the training regime is silk-rand-homo with a backbone of silk-pvgg-4. We then go silk-rand-homo.yaml, which on the 2nd line has silk-default configuration as the default configuartion. Finally, going to silk-default.yaml, we can see the photometric augmentation here, line 20 states that this is the photometric used for IMC (4.3) and ScanNet (4.4.1) which are table 4. and table 5. in the SiLk paper. The ReadMe file stateshere, that the coco-rgb-aug model was used to produce table 4. and 5. results, that is why I believe the default training configuration results in the coco-rgb-aug model.

Of course, the authors can correct me if I pointed you in the wrong direction, hope this helps!

[gleize]

Hi @wangerniuniu,

Thanks @AliYoussef97 for the explanation. Yes, that's exactly right. The default setup is for training the coco-rgb-aug version, which only changes the photometric augmentation.