Customised datasets - combing models now supported?

[korejan]

@Gyx-One Hi, I saw the recent commited pull request and in the description it mentions this:

II.Customise dataset
1.Enable multiple dataset training(eg:MSCOCO and MPII etc,aslong keypoint conversoin is provided)
2.Enable user-add dataset training(eg:user self collected coco-format dataset mixed with official coco dataset)

Does this mean I can could make a new trained model combing body and foot poses (using the coco foot keypoint dataset? I don't even need the full foot pose I would to just like track the tip of the feet. If so, a pointer to docs on how to go about it would be great.

On related note, I wanted to mention this before in the past but have you guys seen the COCO WholeBody dataset?

[One]

Hello! @korejan
Customised dataset means user can combine multiple datasets together as long as they provide the keypoint conversion functions for each dataset they want to combine, so technically speaking, it is possible to combine coco foot keypoint dataset and other datasets (e.g. COCO dataset).

But currently we haven't implement the preprocessing code for coco foot keypoint dataset(which should be used to extract annotation information from the original coco foot keypoint dataset format and put them into the uniform format as the hyperpose.Dataset.mscoco_dataset and hyperpose.Dataset.mpii_dataset).
(we are working on pifpaf now, so more datasets should be implemented sometimes later)

So if you would like to try, you need to write a dataset inheriting the base_dataset and implement functions to extract keypoint annotations from coco foot keypoint dataset following the exact format of mscoco_dataset and mpii_dataset. And then combine the foot dataset object and other dataset object using the multi-dataset combination API. same way to other keypoints dataset. Document updating is on the way! :)

In feature, if time allows, we will implement other datasets(including the coco foot keypoint dataset) and give an example of how to combine them.