Support of MIXAMO for my reddit friend

[AmmarkoV]

Feature request :
https://www.reddit.com/r/gamedev/comments/1trfrkf/comment/ooym19d/?context=1&screen_view_count=2

[zwukong]

--bvh-template ./SAM3DBody-cpp/mixamo.bvh
�[32mFast BVH parser revealed 65 joints / 0 motion fields
�[0m
Number of Values Per Frame: 159
Allocating 636 bytes of memory to hold BVH motions
Frames: 1(1) / Frame Time : 0.0333
mhr_lbs: joints=127 skin=51337 verts=18439 shape_pc=45 face_pc=72
mhr_lbs: loaded scale PCA (mean[68], comps[28x68])
mhr_lbs: loaded hand PCA (mean[54], comps[54x54]) + joint idxs L/R[27]
mhr_lbs: loaded ./onnx/body_model.lbs OK (version 3)
[BVHWriter] mapped 0 BVH joints to MHR; root jID=0
[BVHWriter] failed to map any BVH joint
[main] BVH writer failed to open (continuing without BVH output).

[AmmarkoV]

Can you attach the template file you are using? Is it the one inside the repo ?

[zwukong]

.bvh not supported, just from the main branch

[zwukong]

--bvh-template ./SAM3DBody-cpp/body_mhr.bvh is ok， mixamo.bvh not

[zwukong]

I used the windows package built from other perple ,may be code not the latest

[AmmarkoV]

You can build the latest main yourself. Two documented routes:

1. WSL2 (recommended gets you the full CUDA build, see WSL.md)

• Install gcc-12, CUDA 12.x toolkit (WSL-Ubuntu package), and cuDNN 9.x (cuDNN 8 will not work).
• Set your GPU arch in scripts/build.sh (-DCMAKE_CUDA_ARCHITECTURES=89 for RTX 40xx, 86 for 30xx, 75 for 20xx), then bash scripts/build.sh.
• Before running, fix the WSL library path (Step 6 in WSL.md) or the Windows GPU driver gets picked instead of the WSL one.

2. Native Windows headless (MSVC + CMake, see the "Windows (headless build)" section of README.md)

• MSVC + CMake, OpenCV via vcpkg. CMake auto-fetches the win-x64 ONNX Runtime.
• Builds fast_sam_3dbody_run and offline_sam_3dbody_render. Note the live OpenGL viewer (fast_sam_3dbody_render) is not built on Windows use the BVH/offline output for
  visualization.

Once you're on a current build, --bvh-template ./mixamo.bvh will map (50/65 joints here) and write the BVH.

[zwukong]

it is not very convenient to compile on each update, cuda compile is very complicated and slow i think.
what i want is the new exe and dll shown in the picture. How to make it as fast as possible😄