Open luoww1992 opened 1 week ago
Hi, the goal of this project it to do both detection and regression using a single network. So we do not build on top of yolo detector.
@fabienbaradel about q1, the idx in code is a tuple with 4 tensor: like ( tensor([0], device='cuda:0'), tensor([17], device='cuda:0'), tensor([33], device='cuda:0'), tensor([0], device='cuda:0') ) you say the meaning is the primary keypoint locations, while it doesn't like (x, y). the detail about idx... thank you.
It corresponds to the token location and not (x,y) location in pixel space :)
Is there a method to get the X and Y pixel locations of a joint?
@coolkidmacho this output gives you X/Y pixel locations of SMPL-X joints.
@fabienbaradel What is the orientation of the coordinate system ? the rotvec in humans is euler ?
It is a left-hand coordinate system:
i get human bbox by use yolo8 from image,then i see the func forward code in model.py. i find you use the project: shubham-goel/4D-Humans to get base smpl pose.
q1, in model.py -- class Model() -- func def detection(), what the idx is? the index of label ?
q2. the self.x_attention_head=HPH(...) is used to get head and body base pose by 4D-Humans ?
q3, if q2 is yes, can i use yolo to replace some code ?
q4, if i use the original code from class SMPLTransformerDecoderHead(nn.Module): in 4D-Humans, can i get the same result ? or a litter different ?
q0 -- other : a plan abou onnx or tensorrt for a higher fps ?