Table of contents
2. CelebA (Celebrity Attribute)
1.MS COCO (Common Objects in Context)
DTD（Describable Textures Dataset）
face dataset
1.Helen Face
Dataset introduction:
The Helen Face data set is a data set for face key point detection, which contains 2330 face pictures, covering different postures, expressions, lighting and other conditions. Each picture has a corresponding label, including the position information of the key points of the face. This dataset was developed by Helen's research group in 2011 for training and evaluating face keypoint detection algorithms. Since the Helen Face dataset contains a large number of face images and the key points have been marked, it can be used to train the algorithm to recognize the key points of the face.
In the training image repair algorithm, since the images in the Helen Face dataset cover different postures, expressions, lighting and other conditions, it can be used to train the face key point detection algorithm; it can also be used to train the face repair algorithm, such as training Algorithm to remove face occlusion, blurring and other problems in the image.
Dataset source:
Link: http://www.ifp.illinois.edu/~vuongle2/helen/
Producer: Helen Research Group (a research group composed of the Department of Computer Science and Engineering at the University of Illinois at Urbana-Champaign, USA)
原始文章：A. S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, "CNN features off-the-shelf: an astounding baseline for recognition," in IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2014, pp. 806-813.
Sample example:
2. Celebrity Attribute _ _
Dataset introduction:
The CelebA dataset is a dataset consisting of face images and associated attribute labels. It contains more than 200,000 high-resolution face images from more than 10,000 different celebrities. Each image has 40 different attribute labels, such as gender, age, sideburns, eye color, etc., which can be used to train and evaluate computer vision systems.
The CelebA dataset consists of a large number of high-quality face images, so it can be used to train and evaluate image inpainting algorithms. For example, an algorithm can be trained on the CelebA dataset to inpaint images with occluded faces, or to predict the color and texture of missing parts on the dataset.
Dataset source:
Link: CelebA Dataset
Production institution: Multimedia Laboratory (CUHK) of the Chinese University of Hong Kong
原始文章：Liu, Z., Luo, P., Wang, X., & Tang, X. (2015). Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (pp. 3730-3738)
Sample example :
3. Type A - HQ
Dataset introduction:
The CelebA-HQ dataset is an upgraded version of the CelebA dataset, which is a dataset consisting of high-resolution face images and related attribute labels. It contains over 30,000 high-resolution (1024x1024) face images from over 1,000 different celebrities. Each image has 40 different attribute labels, such as gender, age, sideburns, eye color, etc., which can be used to train and evaluate computer vision systems. The CelebA-HQ dataset is mainly used for high-resolution face generation and super-resolution reconstruction tasks because it provides high-resolution face images and related attribute labels.
Compared with the CelebA dataset, the CelebA-HQ dataset has higher resolution and more samples, so it can be better used for high-resolution face generation and super-resolution reconstruction tasks. Each picture in the CelebA-HQ dataset has a variety of attribute labels, which allows the image restoration algorithm to consider attribute information, such as gender, age, hair color, etc., when repairing the image, so as to better repair the image and get closer to real people The look of the face.
Dataset source
Link: https://github.com/tkarras/progressive_growing_of_gans
Productions: Tero Karras, Samuli Laine, Timo Aila and researchers at NVIDIA
原始文章：Karras, T., Laine, S., & Aila, T. (2018). Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations.
Sample example :
4.FFHQ（Flickr-Faces-HQ）
Dataset introduction:
The FFHQ dataset is a high-quality face image dataset that contains more than 70,000 high-resolution (1024x1024) face images from more than 8,000 different celebrities. Each image has a variety of attribute labels, which can be used to train and evaluate computer vision systems.
The FFHQ dataset is mainly used for the training and evaluation of generative adversarial networks (GANs), because it provides a large number of high-resolution face images and related attribute labels. Compared with the CelebA-HQ dataset, the FFHQ dataset is larger and more diverse, and is more suitable for high-quality face generation tasks.
When training the image inpainting model, using the FFHQ dataset can enable the model to learn the details of high-resolution faces, and attribute labels can be used to ensure that the inpainted image is more in line with the authenticity of the face. For image inpainting algorithms, the FFHQ dataset can be used for tasks such as super-resolution reconstruction, high-resolution face generation and image inpainting, which can better repair facial features such as eyes, nose and mouth.
Dataset source
Link: https://github.com/NVlabs/ffhq-dataset
Production organization： Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, Timo Aila etc. NVIDIA researchers
原始文章：Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., & Aila, T. (2019). A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 4401-4410).
sample example
scene dataset
1.MS COCO (Common Objects in Context)
Dataset introduction:
The MS COCO dataset is a large image dataset used for object detection, semantic segmentation and other computer vision tasks. This dataset contains more than 330,000 images containing more than 250 million object instances.
The main features of the MS COCO dataset are:
1) The amount of data is large, including 80 different object categories.
2) High-quality annotations, each image has multiple annotation boxes, including object category and location information.
For image inpainting algorithms, the MS COCO dataset can be used for the fusion of semantic information, as well as tasks such as object generation and restoration. It helps the model learn to recognize various objects and be able to fix them better. This is very useful for image inpainting algorithms because it enables the model to better recognize and repair objects in the image.
Dataset source
链接：COCO - Common Objects in Context
Production institution: Jointly produced by Microsoft Research Asia, Stanford University, Carnegie Mellon University and other institutions
References: Lin , T.-Y. , Maire , M. , Belongie , S. , Hays , J. , Perona , P. , Ramanan , D. , … & Zitnick , CL (2014, June). Microsoft coconut: Common objects in context. In European conference on computer vision (pp. 740-755). Springer, Cham
Sample example :
2. ImageNet
Dataset introduction:
The ImageNet dataset is a very popular large-scale image dataset mainly used for computer vision tasks such as image classification, object detection and instance segmentation. This dataset contains more than 140,000 images of different categories, divided into 1000 categories in total.
The main features of the ImageNet dataset are:
1) The amount of data is large, including 1000 different object categories.
2) High-quality annotation, each image has multiple annotation boxes, including object categories.
3) The competition will be held every year, and this data set will be used in the competition, which can help researchers evaluate and compare the performance of different algorithms.
For image inpainting algorithms, the ImageNet dataset can be used for the fusion of semantic information, as well as tasks such as object generation and repair. It can help the model learn to recognize various objects and repair them better, making the results of image restoration algorithms more realistic. In addition, it can also be used to evaluate image restoration algorithms, evaluate the performance of restoration algorithms by comparing images before and after restoration, and compare the pros and cons of different algorithms.
Dataset source:
Link : ImageNet
Production Institutions: Stanford University, Caltech, Microsoft Research Asia, University of Washington and others
原始文章：Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009, June). Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (pp. 248-255). IEEE.
Sample example:
3.Places2
Dataset introduction:
The Places2 dataset is a large-scale scene image dataset, which contains a total of 100,000 high-quality scene images.
The main features of the Places2 dataset are:
1) The amount of data is large, including 405 different scene categories.
2) High-quality images, each image is clearly visible.
3) Diversity, the scenes in the dataset include cities, forests, beaches, indoors, etc.
Dataset source
链接：Places2: A Large-Scale Database for Scene Understanding
Produced by: Stanford University and Microsoft Research
原始文章：B. Zhou, H. Zhao, X. Puig, S. Fidler, A. Barriuso and A. Torralba. "Places: A 10 million Image Database for Scene Recognition" IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017
sample example
street view dataset
1.Paris StreetView
Dataset introduction:
The Paris StreetView dataset is a large-scale street view image dataset that contains street view images of Paris, France. A total of more than 1 million high-resolution street view images have been collected, including various buildings, streets, landscapes, etc. in the city.
The Paris StreetView dataset is a large dataset of street view images that can be used to train and evaluate image restoration models in researching image restoration algorithms. Because this data set contains a large number of high-resolution street view images, which may contain image damage, motion blur, noise and other issues. Therefore, this data set can be used to train image inpainting models to better adapt to practical applications Image problem in .
Dataset source
Link: http://opendata.paris.fr/explore/dataset/photos-de-rue-a-paris/
Producer: City of Paris
Original article: (not from the article, is a public dataset)
sample example
2.Cityscapes
Dataset introduction:
The Cityscapes dataset is a high-quality urban street view image dataset, which contains a large number of high-resolution street view images from major European cities (such as Paris, London, Berlin, etc.). This dataset also includes annotation information corresponding to these images, such as roads, buildings, trees, etc.
The Cityscapes dataset is a large dataset of urban street view images that can be used to train and evaluate image inpainting models. Because this dataset contains a large number of high-resolution street view images, which may contain image corruption, motion blur, noise and other issues. Therefore, this dataset can be used to train image inpainting models to better adapt to image problems in practical applications.
In addition, the Cityscapes dataset can also be used for other tasks such as image classification, object detection, semantic segmentation, etc. These tasks can help image inpainting algorithms better understand the structure and semantics of images and improve the performance of image inpainting models. It can also be used in autonomous driving research, urban planning and other fields to help image restoration algorithms better adapt to actual scenarios.
Dataset source
链接：Cityscapes Dataset – Semantic Understanding of Urban Street Scenes
Production organization: Daimler AG, Ford Motor Company, Audi AG, NVIDIA Corporation, and the Technical University of Munich.
原始文章： Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., & Schiele, B. (2016). The Cityscapes Dataset for Semantic Urban Scene Understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
sample example
texture dataset
DTD（Describable Textures Dataset）
Dataset introduction:
The DTD dataset is a large-scale dataset for computer vision and image processing research. It contains over 56,000 images covering over 47 different texture types. Each image has a corresponding text description, which is used to describe the texture features in the image. The DTD dataset is used to train and evaluate image classification and texture analysis algorithms.
The images and corresponding text descriptions in the DTD dataset are used to train image inpainting algorithms. During training, the algorithm learns how to inpaint textures in images based on text descriptions. For example, an algorithm can learn how to repair the stone texture in a corrupted image based on the text description "stone texture". Once trained, the algorithm is able to identify and repair different types of textures in unknown images.
It should be noted that the DTD data set is used for classification and texture analysis algorithms, not for image repair algorithms. The image itself in DTD has not been damaged. When training image repair algorithms, additional processing is required to damage the image Retrain.
Dataset source
Link : http://www.robots.ox.ac.uk/~vgg/data/dtd/.
Producer: The Computer Vision Research Team of the University of Oxford, UK
原始文章："The Describable Textures Dataset (DTD)"E. Cimpoi, M. Maji, S. Mohamed, and I. Kokkinos.Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016.
sample example
building dataset
Façade
Dataset introduction:
The Façade dataset is a very popular dataset for computer vision and deep learning research. It contains a large number of building facade images with different angles, lighting and occlusions. The dimensions of these images are 1024x1024 pixels, and each image has a corresponding label image, which contains the label of the geometry and material type of the building facade. The Façade dataset is widely used to train deep learning models to predict the geometry and material types of building facades.
This dataset has a variety of labels, including:
1) The geometry of building facades, such as windows, doors, walls, etc.
2) The material type of the facade of the building, such as glass, aluminum plate, stone, etc.
3) The lighting state of the building facade, such as shadow, sunlight, etc.
Image inpainting algorithms are often used to restore corrupted or occluded images, such as filling missing parts or removing noise. The images in the Façade dataset have natural occlusion and noise, so it is a good dataset for evaluating the performance of image inpainting algorithms.
For example, researchers can use the Façade dataset to train deep learning models to predict the geometry and material type of building facades, and then use those predictions to repair occluded or missing parts in images. In addition, this dataset can be used to evaluate the performance of different image inpainting algorithms and select the best algorithm.
Dataset source:
Link:
Project on GitHub: https://github.com/shannontian/facade-parsing
Official Website: CMP Facade Database
Dataset sharing platform: https://www.vision.ee.ethz.ch/datasets_extra/facade/
Producer : Czech Technical University in Prague (Czech Technical University)
原始文章：A Dataset for Building Facade Parsing and Its Applications in Automatic Facade Design" by T.Y. Lin, A. Maier, A. Prusa, and J. Kosecka in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016
sample example
