Look, here on the basis of Imagenet a new dataset was collected, in which part of different objects were specified separately: paws/tail/body/head of animals, car body/wheel, etc. In Dataset 24K images from 158 grades of Orininal Imageenet
The authors believe that training on such a dataset will help introduce more to the nature of objects and the interaction of their parts in the ML model. And this, as a result, will help models better to transfer knowledge from domain to domain (Domain Adaptation) and improve models on a small amount of data (Few-Shot Learning) can also improve the quality of segmentation and detection (for example, in the case when you need to detect The object when only its small part is visible).
It sounds good, but there are two nuances:
- Almost no one creates such datasets due to the fact that it is very difficult: you need to spend a lot of time marking each image. Let's hope this work and results of networks trained on Partimagenet will adjust other researchers to create similar datasets. On one such dataset, we will not wait for breakthroughs.
- Ideally, almost every object is sorely divided into parts in different ways. The picture shows that the authors of Partimagenet divided the car into two parts: wheels and a body. But you can divide it differently: highlight the hood, doors, etc. Of course, any division should already somehow help ML models to study, but it is interesting how different divisions will affect the results of the models.