SACD: Subject-Aware Composition Dataset

Guo-Ye Yang¹, Wen-Yang Zhou¹, Yun Cai¹, Song-Hai Zhang¹, Fang-Lue Zhang²

¹Tsinghua University, ²Victoria University of Wellington

Introduction of SACD

We propose a subject-aware composition dataset (SACD) for the task of image cropping for composition. It includes 2,777 images and over 5.2 Million ranking pairs. The labeled cropping windows all have high aesthetic value with a certain focused subject. Some example images of our dataset are shown in Fig. 1. Rather than aimlessly finding cropping windows with good compositions in the entire image, models trained on SACD can give cropping recommendations for a specific subject in the image.

Fig. 1 Subject-aware composition dataset (SACD) examples. (a) some images with multi-subjects in SACD. Original images are tagged with A, and sub-images labeled by professional artists for different subjects are tagged with B1 and B2. (b) some images with single-subject in SACD, meaning of tags are same as (a).

Following is the brief information about the dataset:

• Number of images: 2777;
• Number of cropping windows: >24000;
• Number of ranking pairs: >5.2 Million.

Download

Details about SACD can be found in this paper [link].

Citing

Please cite our paper in your publications if it helps your research: