Way back in 2010, as part of the Pascal Visual Object Challenge, a yearly competition has been held named ImageNet- Large Scale Visual Recognition Challenge which abbreviation is more popular in scientific literature and is known as ILSVRC. It is using a subset of the ImageNet database with roughly 2000 images in each of the 2000 categories. Rough estimates are that contains about 1.3 million training images, 60,000 validation images, and 160,000 testing images. Pretty much to investigate J Luckily, today we have companies that are experienced and with a proven record of finished projects such as Tasq.ai which have experts for this kind of task.
ImageNet is a dataset of over 14 million annotated high-resolution images belonging to roughly 22,000 categories. All those images were collected and labeled by human labelers. That is one more indicator of how important the human workforce is in the development processes of Artificial Intelligence ImageNet is consisting of variable-resolution images, where it is possible to report 2 rates of an error top -5 and top- 1 where the first option (top -5) represents a fraction of the test image where the correct label isn’t in the top 5 labels which model is considered as most probable. There are 3 challenging tasks in ILSVRC:
- Image classification- Ability to predict the classes of the objects on an image
- Single-object localization- It includes Image classification and requires drawing a bounding box around one specific example of each object presented on an image.
- Object detection- Similar to previous, includes Image classification and requires a drawing of a bounding box around every visible object presented on an image.
The improvement pace of ILSVRC was dramatic in the first 5 years, it could be defined as shocking for AI and Computer Vision industry. It was especially visible in a domain of large convolutional neural networks on GPU hardware, which woke the interest in deep learning beyond the data scientist communities, into a mainstream subject.
