liyang
ae28d55f81
Source: RGB-D -> Dyn-HaMR -> L20 retargeting -> FoundationPose -> reference repair -> SPIDER, documented in docs/PIPELINE_LATEST.md and docs/SETUP_AND_WEIGHTS.md. Adds FoundationPose and nvdiffrast upstream snapshots, requirements/pipeline_venv.txt and the FoundationPose weight manifest/downloader. Assets (Git LFS): weights/ (WiLoR detector, HandFlow denoiser, UniDepth-L), FoundationPose checkpoints, HaMeR checkpoint, Dyn-HaMR HMP model and BMC constraints, L20 URDF/meshes, the 20260915_171525 D405 recording and the two box CADs. MANO models are not redistributed (third_party/hamer/_DATA/data/mano/README.txt). Environments, caches and run outputs excluded. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
59 lines
3.1 KiB
Python
59 lines
3.1 KiB
Python
"""Read-only trajectory diagnosis; frame differences are not ground-truth errors."""
|
|
import json
|
|
import sys
|
|
from pathlib import Path
|
|
import numpy as np
|
|
import torch
|
|
from scipy.spatial.transform import Rotation
|
|
ROOT = Path(__file__).resolve().parents[1]
|
|
sys.path.insert(0, str(ROOT/'third_party/Dyn-HaMR/dyn-hamr'))
|
|
from body_model import MANO, run_mano
|
|
from vis.tools import smooth_results
|
|
torch.set_num_threads(4)
|
|
BASE = ROOT/'output/20260915_171525_dynhamr'
|
|
OUT = BASE/'diagnosis'
|
|
OUT.mkdir(exist_ok=True)
|
|
def stats(x):
|
|
x=np.asarray(x)
|
|
return dict(rms=float(np.sqrt(np.mean(x*x))), p95=float(np.percentile(x,95)), maximum=float(np.max(x)))
|
|
def positions(x):
|
|
d=np.linalg.norm(np.diff(x,axis=0),axis=-1)*1000
|
|
dd=np.linalg.norm(np.diff(x,n=2,axis=0),axis=-1)*1000
|
|
return dict(step_mm=stats(d), second_difference_mm=stats(dd), top_step_frames=(np.argsort(d)[-8:][::-1]+1).tolist())
|
|
def rotations(x):
|
|
r=Rotation.from_rotvec(x.reshape(-1,3)).as_matrix().reshape(x.shape[:-1]+(3,3))
|
|
inc=r[1:]@np.swapaxes(r[:-1],-1,-2)
|
|
a=Rotation.from_matrix(inc.reshape(-1,3,3)).magnitude()*180/np.pi
|
|
return stats(a)
|
|
model=MANO(model_path=str(ROOT/'third_party/Dyn-HaMR/_DATA/data/mano'),batch_size=704,pose2rot=True)
|
|
report={}
|
|
saved={}
|
|
for stage in ['smooth_fit','prior']:
|
|
path=sorted((BASE/'optimization'/stage).glob('*world_results.npz'))[-1]
|
|
arr=np.load(path)
|
|
for filtered in [False,True]:
|
|
d={k:torch.from_numpy(arr[k].copy()).float() for k in ['trans','root_orient','pose_body','is_right','betas']}
|
|
if filtered:
|
|
d['root_orient'],d['pose_body'],d['betas'],d['trans']=smooth_results(d['root_orient'],d['pose_body'],d['betas'],d['is_right'],d['trans'])
|
|
with torch.no_grad():
|
|
j=run_mano(model,d['trans'],d['root_orient'],d['pose_body'],d['is_right'],d['betas'])['joints'].numpy()
|
|
key=stage+('_render_filtered' if filtered else '_raw')
|
|
saved[key]=j
|
|
report[key]={}
|
|
for b in range(2):
|
|
side='right' if arr['is_right'][b,0]>.5 else 'left'
|
|
report[key][side]=dict(wrist_world=positions(j[b,:,0]),root_rotation_step_deg=rotations(d['root_orient'][b].numpy()),finger_rotation_step_deg=rotations(d['pose_body'][b].numpy().reshape(352,15,3)))
|
|
report[key]['relative_wrist']=positions(j[1,:,0]-j[0,:,0])
|
|
if stage=='prior':
|
|
R=arr['cam_R'][0]; t=arr['cam_t'][0]
|
|
for b in range(2):
|
|
side='right' if arr['is_right'][b,0]>.5 else 'left'
|
|
report[key][side]['wrist_camera']=positions(np.einsum('tij,tj->ti',R,j[b,:,0])+t)
|
|
cam=np.load(ROOT/'output/20260915_171525/rgbd_camera.npz')
|
|
report['camera_position']=positions(cam['c2w'][:,:3,3])
|
|
report['camera_rotation_step_deg']=rotations(Rotation.from_matrix(cam['c2w'][:,:3,:3]).as_rotvec())
|
|
report['caveat']='Frame differences include true motion; not a causal decomposition or accuracy metric. Render filtering replicated for static views; source camera render excludes translation filtering.'
|
|
np.savez_compressed(OUT/'joints_diagnostic.npz',**saved,time=cam['time'])
|
|
(OUT/'metrics.json').write_text(json.dumps(report,indent=2))
|
|
print(json.dumps(report,indent=2))
|