Datasets used to train NR-IQA metrics

[Stenmarken]

Hello, and thank you for this excellent resource!

I'd like to know the datasets used to train the various NR-IQA metrics. Specifically, I am wondering about DBCNN, Q-Align, and ILNIQE. This page lists the datasets used to train some metrics, but not all of them.

[hakan-emre-gedik]

In their code for Q-Align, they instantiate the model as:

"self.model = AutoModelForCausalLM.from_pretrained(
'q-future/one-align',
trust_remote_code=True,
load_in_4bit=True if dtype == '4bit' else False,
load_in_8bit=True if dtype == '8bit' else False,
torch_dtype=torch.float16 if dtype == 'fp16' else None,
)"

As the one-align model is loaded, as indicated in Q-Align paper, following datasets are used: KonIQ SPAQ KADID LIVE-C AGIQA LIVE CSIQ AVA LSVQtest LSVQ1080P KoNViD MaxWell. Basically, every dataset in IQA research.