You need a dataset for training pose estimation models, but what’s the best choice?
There are public datasets like COCO, MPII, and CrowdPose. Not much, if we compare it to the number of publicly available datasets for different computer vision tasks, like object detection or classification.
The pose estimation problem belongs to a category of rather complex problems. Building a suitable dataset for neural network models is hard. Every joint of every person in the image has to be located and tagged. That’s a mundane and time-consuming task.
The most popular dataset for pose estimation is the COCO dataset. It has around 80 categories of images and around 250,000 instances of people.
If you check some random images from this dataset, you may come across instances that are irrelevant to the problem you are going to solve. Reaching the highest level of precision is desirable in academia, but not always in real-world, production environments.
In the real world, we may be more interested in training models that work well in very specific environments, such as pedestrians, basketball players, gym sessions, etc.

#sklearn #coco-dataset #deep-learning #pose-estimation #pandas