The visualizations seem a bit odd about pose2mesh_release

[hongsukchoi]

Hi @elk-april, thanks for your interest.

Could you provide the visualization of the input 2D joint locations? Inaccurate 2D joint locations may lead to inaccurate 3D mesh output. The input seem to be a truncated image, and this might have affected the 2D joint estimation.

[elk-april]

Thanks for replying.
Actually, the input of 2D joint estimation is a complete image.

[hongsukchoi]

hmm... The 2D joint locations look okay. Did you add pelvis and neck joint locations?
Also the order of 2D joint input should follow this.

self.coco_joints_name = (
            'Nose', 'L_Eye', 'R_Eye', 'L_Ear', 'R_Ear', 'L_Shoulder', 'R_Shoulder', 'L_Elbow', 'R_Elbow', 'L_Wrist',
            'R_Wrist', 'L_Hip', 'R_Hip', 'L_Knee', 'R_Knee', 'L_Ankle', 'R_Ankle', 'Pelvis', 'Neck')

And could you show the visualization with the full 3D mesh output?

[elk-april]

Yeah, I use the center of R_Hip and L_Hip as the locations of Pelvis.
"full 3D mesh output" means all 3D mesh output of the video or complete image?
This is the visualization of one picture:

[hongsukchoi]

Yes, I meant the complete image of the 3D mesh output. I guess you rotated the mesh, right?

I think the overall posture looks right but the limb details look inaccurate.
This may be just from the depth ambiguity, but I think the visual results will be better if you use the pre-trained model from pose2mesh_cocoJ_train_human36_coco_muco. pose2mesh_cocoJ_gt_train_human36_coco is trained with groundtruth input 2D joint locations and subtle errors from the input 2D joint locations in test time might cause larger errors on 3D mesh output.

So I recommend you to test the model with pose2mesh_cocoJ_train_human36_coco_muco and visualize the output.

[elk-april]

Yes, I rotate it.
Actually, I also use "pose2mesh_cocoJ_train_human36_coco_muco" to run the demo, the limb details look inaccurate too...
The upper extremities look better, but the lower extremities are still abnormally bent.

[hongsukchoi]

hmm.

I saw this kind of bent leg outputs when the input images were truncated, but your input look okay.
Could you email me the input image and input 2D joint locations to test myself?
my email: [email protected]

[elk-april]

No problem. Thank you very much!