How to use dataset "posetrack18" for 2D human pose detection

[DuKeM-CSGO]

According to demo, I can use dataset "coco" for detection easily. However, it seems there's not a neck keypoint in the detection result. That's why I want to use another dataset, such as "posetrack18":

python demo/top_down_img_demo.py \
        configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/posetrack18/hrnet_w48_posetrack18_384x288.py \
        hrnet_w48_posetrack18_384x288-5fd6d3ff_20211130.pth \
        --img-root tests/data/posetrack18 \
        --json-file tests/data/posetrack18/annotations/test_posetrack18_val.json \
        --out-img-root vis_results

And I got the result, yeah sure there's the neck keypoint within it. But I found this uses annotation data in test_posetrack18_val.json, which looks like been detected already, written into json and just used. So I can't change the source images at all, even I replace the demo images, I think the detection result is not right.

[ly015]

Hi, thanks for using MMPose. The json file (e.g. test_posetrack18_val.json) is only used to provide information like image paths and ground-truth bboxes, which is needed for a top-down pose model. And the keypoint result is predicted by the model.

To run inference on your own images, you would need to create a json file following the coco or posetrack18 format. The bbox information is required and the keypoint field can be left blank.

[DuKeM-CSGO]

Well thanks for answer, now I write my json and it seems to work. But I removed keys mask_file ignore_regions_x and ignore_regions_y in demo, and I modified bbox_head and bbox in annotations to [0, 0, 1920.0, 1080.0]. Then the result looks different at all(not fault at all though). Requiring the mask file and bboxes data here, how could I get them? Guess they're features of mmdet or others.

(Edit: Got the score of each keypoint. Sorry for the extra reply.)

[jin-s13]

You may try using mmdet for human detection.
https://mmpose.readthedocs.io/en/latest/demo.html#using-mmdet-for-human-bounding-box-detection

Note mask is not necessary, but bboxes are needed.