Support Demoing HigherHRNet

[ThorJonsson]

Hi,

Thank you for your interesting work. I am trying to test HigherHRNet on a collection of videos I've found very challenging for existing human pose estimation models. However when I try to execute the following command:

python demo/video_demo_with_mmdet.py $SCRATCH/mmdetection/configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py $SCRATCH/mmdetection/checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth configs/bottom_up/higherhrnet/coco/higher_hrnet48_coco_512x512.py checkpoints/higher_hrnet48_coco_512x512-60fedcbc_20200712.pth --video-path $SCRATCH/VIBE/sample_video.mp4 --show --out-video-root ./

I get a number of config related issues, the first is that the image size seems to be treated as both an integer as well as a list, I made a temporary fix with a try except statement, but the second issue I'm dealing with now is the lack of a flip_index.

I am not sure what purpose the flip_index has but I've tried making it 0 or [0,1] as I saw in another example and neither of these work.

For a detailed traceback I get

Traceback (most recent call last):
  File "demo/video_demo_with_mmdet.py", line 117, in <module>
    main()
  File "demo/video_demo_with_mmdet.py", line 90, in main
    format='xyxy')
  File "/mmpose/mmpose/apis/inference.py", line 238, in inference_pose_model
    pose = _inference_single_pose_model(model, img_or_path, bbox)
  File "mmpose/mmpose/apis/inference.py", line 194, in _inference_single_pose_model
    return_loss=False, img=data['img'], img_metas=data['img_metas'])
  File ".conda/envs/pytorch150/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in __call__
    result = self.forward(*input, **kwargs)
  File "mmpose/mmpose/models/detectors/bottom_up.py", line 103, in forward
    return self.forward_test(img, img_metas, **kwargs)
  File "mmpose/mmpose/models/detectors/bottom_up.py", line 218, in forward_test
    self.test_cfg['flip_test'])
  File "mmpose/mmpose/core/evaluation/bottom_up_eval.py", line 157, in aggregate_results
    heatmaps[1]) / 2.0 if flip_test else heatmaps[0]

Hope you can help me, I'm really excited about your work and hoping to be able to make contributions in the future.

[innerlee]

Could you post the full error message?

btw I edited the original post to fence the codes

[ThorJonsson]

Traceback (most recent call last):
  File "demo/video_demo_with_mmdet.py", line 110, in <module>
    main()
  File "demo/video_demo_with_mmdet.py", line 90, in main
    format='xyxy')
  File "/mmpose/mmpose/apis/inference.py", line 237, in inference_pose_model
    pose = _inference_single_pose_model(model, img_or_path, bbox)
  File "/mmpose/mmpose/apis/inference.py", line 184, in _inference_single_pose_model
    data = test_pipeline(data)
  File "/mmpose/mmpose/datasets/pipelines/shared_transform.py", line 70, in __call__
    data = t(data)
  File "/mmpose/mmpose/datasets/pipelines/shared_transform.py", line 119, in __call__
    meta[key] = results[key]
KeyError: 'flip_index'

this is what I get if I just run the command.

When I add flip_index to the dictionary that you are parsing to be [0,1] I get the error mentioned above

[ThorJonsson]

I have also verified that the example configuration you give in the docs works on my machine. It's only when I try to modify that command to use higher_hrnet that things go bad.

[innerlee]

Okay, its because the demo was written for top-down models.

Demos for bottom-up will be added soon.

[dmitrysarov]

then maybe rename demo file to point this?

[innerlee]

good idea

[gokcergun]

Okay, its because the demo was written for top-down models.

Demos for bottom-up will be added soon.

Hi, thanks for this great tool! I'm looking for a demo for bottom-up inference. Is it available now?

[jin-s13]

@gokcergun
Please check https://mmpose.readthedocs.io/en/latest/demo.html#d-human-pose-bottom-up-image-demo