V1 Able to detect collision between human and object

[LapisRaider]

1. Can do a AABB check first
2. Afterwards, use VIBE to obtain SMPL parameters and check which vertex or keypoint is the nearest to the object
3. If there's no collision, do not attach

With this in the second iteration we just need to change the attachment offset code and the collision offset instead of fully relying on the detector

[LapisRaider]

Rendering\lib\data_utils\kp_utils.py and Rendering\lib\models\smpl.py contains the mapping

[LapisRaider]

We are using SPIN models, for the joint mappings. Thus, we using get_spin_joint_names()

[LapisRaider]

Difference between SMPL and SPIN

get_smpl_joint_names and get_spin_joint_names is different

SPIN

• There are 49 joints in total https://github.com/mkocabas/VIBE/blob/master/lib/data_utils/kp_utils.py#L212
• Combination of openpose keypoints and a joint regressor trained to match Human3.6M dataset (motion capture dataset) joint definitions.
• https://github.com/nkolot/SPIN/issues

SMPL

• SMPL pose parameters have 72 elements, represent rotations of 24 jopints in axis-angle format
• SMPL has 24 joints + 21 joints compatible with 2D keypoint datasets.

[LapisRaider]

Can use joints2d_img_coord from the smpl pkl output file Array storing for each frame 49 keypoint, most likely SPIN's model 49 joints

[LapisRaider]

Pose data uses fbx_output.py bone_name_from_index

[LapisRaider]

SMPL joints - 23: https://www.researchgate.net/figure/Layout-of-23-joints-in-the-SMPL-models_fig2_351179264