How were the annotated images in person_keypoints_minitrain2017.json selected?

[automate11]

Thank you for the contributions your research has made, and I appreciate you taking the time to look into my question.

Regarding the “Keypoint Detection” section in the code’s homepage under the “download” chapter, where you shared the COCO-style JSON file for person_keypoints_minitrain2017.json, could you explain how this file was generated?

I cross-checked and noticed that some images in the coco_minitrain_25k.zip downloaded from Kaggle, specifically in the images folder, do not have corresponding annotations in this JSON file. For example, the image 000000000692.jpg in the minicoco dataset has no annotation in the JSON file.

Does this mean the JSON file only includes a subset of the images with person annotations?

How was this subset of data selected?

Looking forward to your response.😊😊😊

[automate11]

The file_name should be a unique identifier for each image, so I searched by file_name and found that the result shows 15,000 images. This implies that the keypoint JSON file only contains annotation information for 15,000 images, missing annotations for many other images containing people. Why is that?

[automate11]

And when I changed the dataset’s JSON file to person_keypoints_minitrain2017.json and the image dataset to minicoco, I encountered an issue where the image names listed in the JSON file couldn’t be found in minicoco.

However, one way to solve this problem is to use the original train image dataset from COCO instead of minicoco.

With this setup, the program can locate the corresponding images in the train dataset based on the names listed in the JSON file, and it successfully loads the annotated images.

I’m wondering if this is, in fact, the correct way to proceed.

Originally, I thought that person_keypoints_minitrain2017.json could be used in conjunction with the minicoco dataset.

I hope the author can clarify the correct usage whenever possible, and I would be very grateful for any insights.😊😊😊

[giddyyupp]

Hello, minitrain for object detection and pose estimation are two separate datasets. For pose estimation, we created a new subset of COCO using the images with person label. So it is a subset of COCO's person_keypoints_train2017.json and different than the instances_minitrain2015.json

hope this clarifies your questions.