On April 30, Google issued a document on its official blog that it will open the Images V4 database and open the CCTV 2018 Open Image Challenge at the same time.
Here is the compiled information provided by the Google on the open Images on the blog.
In 2016, we released a dataset, Open Images, that contains approximately 9 million images and labels thousands of object categories. After the release, we have been working hard to update and improve data sets to provide a useful resource for the computer vision community to develop new models.
Today, we are pleased to announce the opening of Open Images V4 , which contains 15.4 million bounding boxes for 600 categories on 1.9 million images, which is the largest existing dataset with object location annotations . Most of these boxes are manually drawn by professional commentators, ensuring their accuracy and consistency. In addition, these images are very diverse and often contain complex scenes of multiple objects (8 images per image on average).
At the same time, we will also announce the launch of the Open Images Challenge , which will be a new object detection challenge at the 2018 Computer Vision Europe Conference ( ECCV 2018 ). The Open Images Challenge will follow the traditions of PASCAL VOC , ImageNet and COCO , but its scale will be unprecedented.
The Open Images Challenge will be unique in these areas:
- There are 1.7 million training images, including 500 categories and 12.2 million border annotations;
- Compared to previous detection challenges, there will be a wider category, including new objects such as “fedora”, “snowman”, etc.
- In addition to mainstream object detection, this test will also include visual relationship detection when detecting object pairs, such as “woman playing guitar”.
The training data set is now available; a test set containing 100,000 images will be published on Kaggle on July 1, 2018. The deadline for the submission of the Challenge is September 1, 2018.
We hope that the larger training set will stimulate research on more complex detection models that will exceed the performance of the current state-of-the-art; on the other hand, we hope that the 500 categories can evaluate different probes more accurately. In which areas the device performs better. In addition, having a large number of images with multiple object annotations can help you explore visual relationship detection. This is a hot topic and has more and more sub-communities.
In addition to the above, Open Images V4 also contains 30.10 million manually verified pictures for 19794 class image-level tags. Of course, these tags are not part of the challenge, of which 5.5 million image-level tags were generated by crowdsource.google.com from thousands of users around the world