Computer Vision systems use ‘Object Recognition’ tech to identify an object in videos or images, like any household item. Many companies including Amazon, Google and Microsoft have their own object recognition AI, which are improving with time.
Google Lens is one of the examples of object recognition tech. Blind people are facilitated through it as the same technique is used by Facebook to add ‘Alternative Text’ automatically in photos description.
Facebook AI team published a study with a whole new side of the story.
The study revealed that object recognition is more efficient for people who earn more. Like households with $3500 per month can get better results as compared to a household with a monthly income of $50.
Object recognition systems by Amazon, Clarifai, Facebook, Google, IBM and Microsoft were used in the study. The study does not mention individual results, however, it presented an aggregated result that these systems work 20 percent better for households with a higher income than the poor ones.
Open source data set, DollarStreet was used to test the systems. The data was based on items of 264 homes belonging to 50 countries.
When it comes to geographical distribution, object recognition works better for the US and Europe’s objects as compared to Africa and Asia.
The major reason behind this could be the data used for training of the system. Researchers explained that the data used for training was taken from ImageNet and COCO. These datasets have much of the objects belonging to the US and Europe instead of Africa or Asia.
In addition to it, these systems are also trained by using queries which are mostly in English and thus, these research queries may also have an impact on the recognition results of the objects. Language or vocabulary difference can have a huge impact when searching any query.
Read next: Google's DeepMind AI System can now Successfully Compete Against Humans in Video Games
Featured illustration: Freepik / Supha
Google Lens is one of the examples of object recognition tech. Blind people are facilitated through it as the same technique is used by Facebook to add ‘Alternative Text’ automatically in photos description.
Facebook AI team published a study with a whole new side of the story.
The study revealed that object recognition is more efficient for people who earn more. Like households with $3500 per month can get better results as compared to a household with a monthly income of $50.
Object recognition systems by Amazon, Clarifai, Facebook, Google, IBM and Microsoft were used in the study. The study does not mention individual results, however, it presented an aggregated result that these systems work 20 percent better for households with a higher income than the poor ones.
Open source data set, DollarStreet was used to test the systems. The data was based on items of 264 homes belonging to 50 countries.
- Also read: Microsoft creates a text-to-speech Artificial Intelligence that can generate realistic speech
When it comes to geographical distribution, object recognition works better for the US and Europe’s objects as compared to Africa and Asia.
The major reason behind this could be the data used for training of the system. Researchers explained that the data used for training was taken from ImageNet and COCO. These datasets have much of the objects belonging to the US and Europe instead of Africa or Asia.
In addition to it, these systems are also trained by using queries which are mostly in English and thus, these research queries may also have an impact on the recognition results of the objects. Language or vocabulary difference can have a huge impact when searching any query.
"The development of object recognition models that work for everyone will likely require the development of training algorithms that can learn new visual classes from few examples and that are less susceptible to statistical variations in training data.", explained study.AI systems can be improved by ending or at least decreasing the biases. In this regard, researchers are planning to use different hashtags of other languages than English for the training of a system. Also, assigning the location-information to visual-datasets can improve AI as well.
Read next: Google's DeepMind AI System can now Successfully Compete Against Humans in Video Games
Featured illustration: Freepik / Supha
