图像裁剪资料汇总

最新推荐文章于 2023-04-26 23:45:00 发布

原创

最新推荐文章于 2023-04-26 23:45:00 发布 · 1.3k 阅读

2 ·

CC 4.0 BY-SA版权

文章标签：

#图像裁剪 #合集 #资料汇总

本文介绍了多个用于图像裁剪的数据集，包括CPC、GAICD等，以及相关的评价标准如SRCC、AccK/NIoU和BDE。此外，还讨论了一系列研究论文，这些论文涉及了美学意识的强化学习方法和深度网络在自动图像裁剪中的应用。

数据集

densely annotated (multiple crops in each image are annotated)

CPC (2018) [link] （10797 images， each image has 24 crops）
GAICD (2019) [link] （1236 images，86 crops for each image）

sparsely annotated (only the best crop in each image is annotated)

KU-PCP（2018） [Paper] [Download link].
ICDB/MSR-ICD (2013) [link] 950 images
FLMS/HCDB (2014) [Download Images] [Download Crops] （500 images）
FCDB (2017) [link] （1743 images, 1395 images are used for training and 348 for testing）

MSR-ICD：950 images which are originally from an image aesthetics assessment database（“Content-based photo quality assessment）。image categories, including animal, architecture, human, landscape, night, plant and man-made objects。Each image is carefully cropped by three expert photographers.

CPC: The CPC dataset (Wei et al. 2018) is the first densely annotated image cropping dataset. It contains 10797 images, and each image has 24 crops with multiple scores annotated by Amazon Mechanical Turk workers. We use the average annotated score for each crop as ground truth.

GAICD: The GAICD dataset (Zeng et al. 2019) is also a densely annotated dataset, but each image has more anno- tated crops than the CPC dataset. It has 1236 images and about 86 crops for each image.

FCDB&FLMS: The FCDB (Chen et al. 2017a) and FLMS (Fang et al. 2014) datasets only have annotations for the best crop in each image, which are given by ten annotators. FCDB contains 358 images, and FLMS contains 500 images. These two datasets are only used as the test set in the image cropping task