SonnetMakingNewRobloxRig / finalize_a.py
DenSnoychik's picture
Upload 33 files
3a6f70b verified
Raw History Blame Contribute Delete
9.56 kB
import pickle, json, numpy as np, scipy.sparse as sp, scipy.sparse.csgraph as cg, os
from scipy.ndimage import binary_closing
from common import *
from fit import Surf
ROOT='..'
res=pickle.load(open('fit_result.pkl','rb')); ref=ref_world()
def clean(V,F,maxedge=None):
if maxedge:
el=np.max([np.linalg.norm(V[F[:,i]]-V[F[:,(i+1)%3]],axis=1) for i in range(3)],axis=0); F=F[el<maxedge]
n=len(V); E=np.vstack([F[:,[0,1]],F[:,[1,2]],F[:,[2,0]]])
A=sp.coo_matrix((np.ones(len(E)),(E[:,0],E[:,1])),shape=(n,n)); k,lab=cg.connected_components(A,directed=False)
fl=lab[F[:,0]]; cnt=np.bincount(fl,minlength=k)
keep=cnt[fl]>=max(0.04*cnt.max(),30); F=F[keep]
used=np.unique(F); rm=-np.ones(n,int); rm[used]=np.arange(len(used)); return V[used],rm[F]
def vnormals(V,F):
N=np.zeros_like(V); e=[V[F[:,1]]-V[F[:,0]],V[F[:,2]]-V[F[:,1]],V[F[:,0]]-V[F[:,2]]]
fn=np.cross(e[0],-e[2])
for k in range(3):
a=-e[(k+2)%3]; b=e[k]
ang=np.arccos(np.clip((a*b).sum(1)/np.maximum(np.linalg.norm(a,axis=1)*np.linalg.norm(b,axis=1),1e-12),-1,1))
nn=fn/np.maximum(np.linalg.norm(fn,axis=1,keepdims=True),1e-12)
np.add.at(N,F[:,k],nn*ang[:,None])
return N/np.maximum(np.linalg.norm(N,axis=1,keepdims=True),1e-12)
final={}
for n in PART_ORDER:
V,F=clean(*res[n],maxedge=(0.28 if n.endswith(('Hand','Foot')) else 0.5)); final[n]=(V,F)
pickle.dump(final,open('final_parts.pkl','wb'))
# ---- OBJ ----
lines=['# rugc_reference_boy.obj - R15 mannequin for Roblox UGC accessory fitting','# Units: Roblox studs, Y-up, character front faces -Z (R15 Fixed rest pose, root-relative world coordinates)',
'# Base mesh: MakeHuman hm08 base mesh (CC0), deformed part-by-part onto the ROBLOX Boy (Bundle 109) proportions. See LICENSE_AND_SOURCE.md','mtllib rugc_reference_boy.mtl']
off=0; vtxt=[];ntxt=[];ftxt=[];tot_v=tot_f=0
for n in PART_ORDER:
V,F=final[n]; N=vnormals(V,F)
ftxt.append(f'o {n}'); ftxt.append('usemtl Mannequin'); ftxt.append('s 1')
vtxt+= [f'v {x:.4f} {y:.4f} {z:.4f}' for x,y,z in V]
ntxt+= [f'vn {x:.3f} {y:.3f} {z:.3f}' for x,y,z in N]
ftxt+= [f'f {a+off+1}//{a+off+1} {b+off+1}//{b+off+1} {c+off+1}//{c+off+1}' for a,b,c in F]
off+=len(V); tot_v+=len(V); tot_f+=len(F)
# write: all v, vn first then parts
out=lines+vtxt+ntxt+ftxt
open(f'{ROOT}/rugc_reference_boy.obj','w').write('\n'.join(out)+'\n')
open(f'{ROOT}/rugc_reference_boy.mtl','w').write('# rugc_reference_boy.mtl\nnewmtl Mannequin\nKa 0.35 0.30 0.27\nKd 0.82 0.70 0.58\nKs 0.05 0.05 0.05\nNs 10\nd 1.0\nillum 2\n')
print('obj',tot_v,tot_f,os.path.getsize(f'{ROOT}/rugc_reference_boy.obj')/1e6,'MB')
# ---- attachments ----
def euler(R):
R=np.array(R); sy=np.sqrt(R[0,0]**2+R[1,0]**2)
x=np.arctan2(R[2,1],R[2,2]); y=np.arctan2(-R[2,0],sy); z=np.arctan2(R[1,0],R[0,0]) if sy>1e-6 else 0
return [round(float(np.degrees(v)),4)+0.0 for v in (x,y,z)]
att={'schema':'roblox-r15-attachments-v1','units':'studs','pose':'R15 Fixed rest pose (arms/legs as in ROBLOX Boy bundle 109)','axes':'Y up, character front = -Z, character Right = +X',
'note':'Attachment CFrames are the real values stored in the Roblox assets of ROBLOX Boy: bundle 109 package assets 376532000/376530220/376531012/376531300/376531703 (R15Fixed folder) and the Roblox Classic Head asset 2432102561 (Vector3Value attachments). position/orientation are LOCAL to the part origin (partCFrame).','parts':{}}
awx={}
for n in PART_ORDER:
d=ref[n]; t=d['t']; R=d['R']
pa={'partCFrame':{'position':[round(float(x),5) for x in t],'rotationMatrix':np.round(R,6).tolist()},'attachments':{}}
for a,(Ra,p) in d['att'].items():
Ra=np.round(np.array(Ra),6)+0.0
wp=R@p+t; awx[(n,a)]=wp
pa['attachments'][a]={'position':[round(float(x),5) for x in p],'orientationEulerXYZdeg':euler(Ra),'rotationMatrix':Ra.tolist(),'worldPositionInRestPose':[round(float(x),5) for x in wp]}
att['parts'][n]=pa
json.dump(att,open(f'{ROOT}/mannequin_attachments.json','w'),indent=1)
# ---- metrics ----
def merge(parts,src):
Vs=[];Fs=[];o=0
for n in parts:
V,F=src[n]; Vs.append(V); Fs.append(F+o); o+=len(V)
return np.vstack(Vs),np.vstack(Fs)
refp={n:(ref[n]['V'],ref[n]['F']) for n in PART_ORDER}
def sample(V,F,n,seed=0):
A=V[F[:,0]];B=V[F[:,1]];C=V[F[:,2]]; ar=np.linalg.norm(np.cross(B-A,C-A),axis=1)/2
rng=np.random.default_rng(seed); ids=rng.choice(len(F),n,p=ar/ar.sum()); u=rng.random((n,2)); m=u.sum(1)>1; u[m]=1-u[m]
return A[ids]+u[:,:1]*(B-A)[ids]+u[:,1:]*(C-A)[ids]
from scipy.spatial import cKDTree
VIEWS6={'front':(0,1,-1),'back':(0,1,1),'left':(2,1,-1),'right':(2,1,1),'top':(0,2,1),'bottom':(0,2,-1)}
def mask(P,view,lo,hi,res=512):
# orthographic silhouette via splat
ax={'front':(0,1,1),'back':(0,1,-1),'left':(2,1,1),'right':(2,1,-1),'top':(0,2,-1),'bottom':(0,2,1)}[view]
u=P[:,ax[0]]*ax[2]; v=P[:,ax[1]]
# square pixels from global bounds
lo_,hi_=lo,hi
ix=((u-lo_)/(hi_-lo_)*(res-1)).astype(int); iy=((v-lo_)/(hi_-lo_)*(res-1)).astype(int)
ok=(ix>=0)&(ix<res)&(iy>=0)&(iy<res)
M=np.zeros((res,res),bool); M[iy[ok],ix[ok]]=True
return binary_closing(M,iterations=2)
def iou(a,b): return float((a&b).sum()/max((a|b).sum(),1))
metrics={'parts':{}}
LO,HI=-3.4,3.4
names=PART_ORDER
for n in names:
mv,mf=final[n]; rv,rf=refp[n]
pm=sample(mv,mf,120000,1); pr=sample(rv,rf,120000,2)
t1=cKDTree(pr); t2=cKDTree(pm); d1=t1.query(pm)[0]; d2=t2.query(pr)[0]
ious={v:round(iou(mask(pm,v,LO,HI),mask(pr,v,LO,HI)),4) for v in VIEWS6}
metrics['parts'][n]={'silhouetteIoU':ious,'meanSilhouetteIoU':round(float(np.mean(list(ious.values()))),4),
'chamfer_mean_studs':round(float((d1.mean()+d2.mean())/2),4),'surface_dist_95pct_studs':round(float(max(np.percentile(d1,95),np.percentile(d2,95))),4),'hausdorff_studs':round(float(max(d1.max(),d2.max())),4),
'bboxMine':[np.round(mv.min(0),3).tolist(),np.round(mv.max(0),3).tolist()],'bboxRef':[np.round(rv.min(0),3).tolist(),np.round(rv.max(0),3).tolist()],
'vertices':int(len(mv)),'triangles':int(len(mf))}
mV,mF=merge(names,final); rV,rF=merge(names,refp)
pm=sample(mV,mF,500000,3); pr=sample(rV,rF,500000,4)
metrics['whole']={'silhouetteIoU':{v:round(iou(mask(pm,v,LO,HI),mask(pr,v,LO,HI)),4) for v in VIEWS6}}
metrics['whole']['meanSilhouetteIoU']=round(float(np.mean(list(metrics['whole']['silhouetteIoU'].values()))),4)
d1=cKDTree(pr).query(pm)[0]; d2=cKDTree(pm).query(pr)[0]
metrics['whole'].update({'chamfer_mean_studs':round(float((d1.mean()+d2.mean())/2),4),'surface_dist_95pct_studs':round(float(max(np.percentile(d1,95),np.percentile(d2,95))),4),'hausdorff_studs':round(float(max(d1.max(),d2.max())),4),
'heightMine':round(float(mV[:,1].max()-mV[:,1].min()),4),'heightRef':round(float(rV[:,1].max()-rV[:,1].min()),4),'bboxMine':[np.round(mV.min(0),3).tolist(),np.round(mV.max(0),3).tolist()],'bboxRef':[np.round(rV.min(0),3).tolist(),np.round(rV.max(0),3).tolist()],'vertices':int(len(mV)),'triangles':int(len(mF))})
# ---- attachment fit zones ----
def ray_first(V,F,o,d):
A=V[F[:,0]];e1=V[F[:,1]]-A;e2=V[F[:,2]]-A
h=np.cross(d,e2); a=(e1*h).sum(1); ok=abs(a)>1e-12; f=np.where(ok,1/np.where(ok,a,1),0)
s=o-A; u=f*(s*h).sum(1); q=np.cross(s,e1); v=f*(q@d); t=f*(e2*q).sum(1)
m=ok&(u>=0)&(v>=0)&(u+v<=1)&(t>1e-5)
return float(t[m].min()) if m.any() else None
RAYS={'BodyFrontAttachment':(0,0,-1),'BodyBackAttachment':(0,0,1),'WaistFrontAttachment':(0,0,-1),'WaistBackAttachment':(0,0,1),'FaceFrontAttachment':(0,0,-1),'FaceCenterAttachment':(0,0,-1),
'HatAttachment':(0,1,0),'HairAttachment':(0,1,0),'NeckAttachment':(0,1,0),'LeftCollarAttachment':(0,1,0),'RightCollarAttachment':(0,1,0),'LeftShoulderAttachment':(0,1,0),'RightShoulderAttachment':(0,1,0),'LeftGripAttachment':(0,-1,0),'RightGripAttachment':(0,-1,0),'LeftFootAttachment':(0,-1,0),'RightFootAttachment':(0,-1,0)}
mS={n:Surf(*final[n],n=60000) for n in names}; rS={n:Surf(*refp[n],n=60000) for n in names}
fit=[]
for (pn,an),wp in awx.items():
if an not in RAYS: continue
d=np.array(RAYS[an],float)
# ray against the whole body (all parts) so overlapping parts count
def sgn(V,F):
tp=ray_first(V,F,wp,d); tn=ray_first(V,F,wp,-d)
c=[x for x in (tp,-tn if tn is not None else None) if x is not None]
return tp if tp is not None else (-tn if tn is not None else None)
hm=sgn(mV,mF); hr=sgn(rV,rF)
nm=min(np.linalg.norm(mS[q].project(wp[None])[0]-wp) for q in names); nr=min(np.linalg.norm(rS[q].project(wp[None])[0]-wp) for q in names)
fit.append({'part':pn,'attachment':an,'worldPos':np.round(wp,4).tolist(),'rayDir':list(RAYS[an]),'signedAxisDistanceToSurface_ref':None if hr is None else round(hr,4),'signedAxisDistanceToSurface_mine':None if hm is None else round(hm,4),
'axisDistanceDelta':None if (hr is None or hm is None) else round(hm-hr,4),'nearestSurface_ref':round(float(nr),4),'nearestSurface_mine':round(float(nm),4)})
metrics['attachmentFit']=fit
json.dump(metrics,open('metrics_tmp.json','w'),indent=1)
print(json.dumps({k:(v['meanSilhouetteIoU'],v['chamfer_mean_studs'],v['surface_dist_95pct_studs']) for k,v in metrics['parts'].items()},indent=0))
print(metrics['whole']['silhouetteIoU'],metrics['whole']['chamfer_mean_studs'],metrics['whole']['surface_dist_95pct_studs'],metrics['whole']['hausdorff_studs'])
for f in fit: print(f['part'],f['attachment'],f['signedAxisDistanceToSurface_ref'],f['signedAxisDistanceToSurface_mine'],f['axisDistanceDelta'],f['nearestSurface_ref'],f['nearestSurface_mine'])