Potential Alternatives

Best options are the ones with less than 6M parameters and 0.5 GFLOPs.

Needed Adjustments to Replace the Backbone

To match the resolution of VGG-19 upsampling or downsampling might be needed. While to match the channels a 1 * 1 conv layer might be needed.

Task-Specific Adaptation

• OpenPose’s two branches (confidence maps and PAFs) rely on task-specific features (e.g., body part locations and limb orientations).
• A new backbone (even pre-trained on ImageNet) won’t naturally produce these features unless fine-tuned for pose estimation.