TADAC:Text Annotated Distortion, Appearance and Content

Citation Author(s):
Submitted by:
Tianhao Gu
Last updated:
Wed, 07/10/2024 - 22:57
Data Format:
Research Article Link:
0 ratings - Please login to submit your rating.


In order to train the joint contrastive representation learning module, we constructe a large Text Annotated Distortion, Appearance and Content (TADAC) image database. We collect images with both synthetic distortions and authentic distortions from multiple image databases.Each image in TADAC is annotated with three types of textual descriptions indicating the semantic contents, the distortion characteristics and the appearance properties respectively.For image content description text,we adpot the textual template format of "A photo of a(n) {s}",where {s} is given by the scene label or object label of the image.For synthetic distorction images,we design a list of phrases describing distortion types and another list of adjectives, adverbs, and quantifiers describing the degree of distortions.For authentic distorction, we design appearance characteristics based on 4 quality relevant aspects:brightness,contrast,sharpness and colorfulness.We constructe a database containing images of diverse contents, a variety of distortions and rich appearances.Importantly, these images are annotated with texts describing their semantic contents, distortion charactistics and appearance propertities.These texts can represent quality relevant high level knowledge.The database can be used in other contexts and will be particularly useful for developing IQA applications



Prepare corresponding dataset

The TADAC is developed based on existing dataset. It is essential to prepare the corresponding dataset beforehand.

Synthetic distortion:KADIS
Authetic distortion:AVA,COCO,VOC,Places365,CERTH-Blur

Download the csv file

The "ugc.caption.csv" is the captions of authetic distorction images.The "sys_caption.csv" is the captoins of synthetic distorction images.

Both files are the same structure, the first column is the image name,the second column is the image caption, and the third column is the image distortion label.