Data Problem In Computer Vision Field

When a problem can be described by a limited set of rules, computers perform very well and usually surpass human performance in terms of speed and accuracy. However, when the problem is too hard to be formulated, computers fail dramatically. Human reasoning and intelligence are still very far from being understood or formulated. The vast research done in this field still only scratches the surface. Visual perception is one of the most complex tasks that the human brain performs accurately and efficiently. As visual perception is important for human daily activities, it is critical for the machine (agent) to perceive the environment, reason, plan, and then interact to achieve its goal.

The machine could be as simple as an Optical Character Recognition (OCR) program or as complex as an autonomous car or robot on another planet. Failure of such agents could lead to deaths or injuries, damage to the environment or properties, failure of missions, and loss of millions of dollars.

The recent great success of Artificial Neural Networks (ANN) in solving complex problems motivated many researchers to apply it to machine perception problems too.

In parallel to this, the great advancement in chip design, microelectronics, and the introduction of General-Purpose Graphics Processor Architectures like the General Purpose Graphics Processing Unit (GPGPU), facilitated training deep neural networks with millions of parameters.

These deep ANNs with their high degree of non-linearity, help in approximating the real functions or phenomena behind complex vision problems (like semantic segmentation, instance segmentation, and object recognition) in a much more accurate way. Unfortunately, training these deep learning models requires a great amount of data together with their corresponding annotations or ground-truths. Finding, collecting, and annotating suitable data is cumbersome, time-consuming, error-prone, expensive, and subject to privacy issues.

Perhaps, the lack of diverse, high quality, and precisely labeled data can be attributed to the previously mentioned reasons. Unfortunately, these factors cause many major data quality issues in the field of computer vision and they clearly present an obstruction toward the aim of optimal performance computer vision models in practice.