Test-Time-Adaption

This repo contains code from two repositories - 1. TTA for Blind Image Quality Assessment and 2.Depth-Anything
The two repositories have been merged only for MetaIQA and for KONIQ dataset inside TTA folder (TTA/TTA-IQA/MetaIQA)

Datasets

Detailed description and link of datasets is given in the TTA-IQA repo. Each dataset needs to be prepared according to the folders.py inside TTA folder (TTA/TTA-IQA/MetaIQA/folders.py)

Pre-trained models

Both the IQA and Depth-Anything pre-trained models are available in their respective repositories. Note torchhub folder form Depth-Anything repo.

Code files

Every TTA model's folder has a ALL_EXPT.py file that lists the experiments and arguments the author pases while testing. MetaIQA folder has been updated by integrating the above mentioned repositories. The following briefly explains each code file :

• dataloader.py and folders.py : prepares the dataset by extracting the MOS code, image paths and defined the pre-processing. NOTE: Dataset needs to be prepared according to how it is being used in the folders.py file. For example, the datasets may have images, MOS code and other information stored in separate files which maybe requied to be extracted and filled in a single .csv file which must be named exactly as given in folders.py file.
• sam.py : this file has been taken from this repo. Simply copy this file in each model's folder to use it.
• ttt.py : this is the main file of MetaIQA folder. It includes all the arguments and some arguments have been added such as each loss' weights and choice of optimizer (either Adam of SAM)
• depth-anything : this folder is taken from Depth-Anything repo and is the only file which is required to run their model. It has been imported in ttt.py file .