Strategy for single object keypoint detection without bounding boxes during inference

[NCTU-VRDL]

Hi!

Thanks for the great tools for keypoint detection. I have a custom dataset, and each image contains only a single object with 19 key points. I have successfully trained the top-down model using this config.

However, in my case, I don't have the ground truth boxes for inference on the new image. For training, I just used the enclosed box that covers all points.

Since I don't want to train another object detection model, should I use the Bottom-up method instead? Any suggestions for this kind of dataset are very welcome. Thanks

[ly015]

Yes, a top-down pose model like the one you used would need object bounding boxes during inference. If a detection model is undesired in your case, you could try the bottom-up method, like associative embedding.

[jin-s13]

If the object is roughly at the center of your image, you may choose to train a model that directly takes the whole image as the input.

[NCTU-VRDL]

Hi @ly015
Thanks for the AE method. Do you have any suggestions for the configuration? It looks like I can set the max_num_people=1 in test_cfg. Anything else?

Hi @jin-s13
Thanks a lot for your response. Unfortunately, the location of the object varies a lot in each image. I will try to use the whole image as the box instead of the enclosed one.