Some suggestion
#3
by
Hiigaraxiao
- opened
Thank you for providing such an innovative model and approach. Based on actual usage, I sincerely offer some suggestions and look forward to your improvements in the model and updates:
- I found that many seemingly normal hands are still not being detected. I feel that this area could benefit from adding some bad hand training data to improve recognition rates.
- When two hands overlap, one hand is often not detected, which causes the non-overlapping hand to become deformed after redrawing.
- The redrawing area may appear incongruous on the original image, especially with complex backgrounds. It feels like feathering could be added to address this issue.
@Hiigaraxiao I'm just an uploader, not the actual author itself but I'll still answer ngl
- The workflow of HandRefiner is MediaPipe segments hands -> MeshGraphormer reconstructs and corrects the hands, outputs MANO mesh -> Ray-tracing with Trimesh to get depth map. It's possible to replace MediaPipe with a YOLO model trained on bad hands. ADetailer uses one iirc
- Again, MediaPipe fault here
- I've seen people using GrowMaskWithBlur node by Kijai to make good masks from the hand depth map
You can post this issue on the original repo if you want further diccussion
hr16
changed discussion status to
closed
could it be used to improve it?
Offical pytorch implementation of "HandOccNet: Occlusion-Robust 3D Hand Mesh Estimation Network", CVPR 2022
https://github.com/namepllet/HandOccNet.git