File size: 13,760 Bytes
1ecb99d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
#!/usr/bin/env python3
import argparse
import json
import shutil
import subprocess
import time
from pathlib import Path

import cv2
import numpy as np
import torch
import triton as tr
import triton.language as tl
import triton.language.extra.cuda.libdevice as lib
from scipy.spatial.transform import Rotation, Slerp


@tr.jit
def field(x, y, z, ball, rod):
    # Nodes at integer grid coordinates; no finite array or moving window.
    qx = x - tl.floor(x + 0.5)
    qy = y - tl.floor(y + 0.5)
    qz = z - tl.floor(z + 0.5)
    sx, sy, sz = qx*qx, qy*qy, qz*qz
    ds = tl.sqrt(sx + sy + sz) - ball
    dxy = tl.sqrt(sx + sy) - rod
    dxz = tl.sqrt(sx + sz) - rod
    dyz = tl.sqrt(sy + sz) - rod
    return tl.minimum(ds, tl.minimum(dxy, tl.minimum(dxz, dyz)))


@tr.jit
def render_kernel(CAM, OUT, DEPTH, W: tl.constexpr, H: tl.constexpr,
                  BALL: tl.constexpr, ROD: tl.constexpr, FAR: tl.constexpr,
                  STEPS: tl.constexpr, BLOCK: tl.constexpr,
                  DIRECTION_RGB: tl.constexpr, BACKGROUND_ONLY: tl.constexpr):
    i = tl.program_id(0)*BLOCK + tl.arange(0, BLOCK)
    valid = i < W*H
    px = (i % W - tl.load(CAM+2)) / tl.load(CAM)
    py = (i // W - tl.load(CAM+3)) / tl.load(CAM+1)
    inv = tl.rsqrt(px*px + py*py + 1.)
    px, py, pz = px*inv, py*inv, inv
    dx = tl.load(CAM+4)*px + tl.load(CAM+5)*py + tl.load(CAM+6)*pz
    dy = tl.load(CAM+7)*px + tl.load(CAM+8)*py + tl.load(CAM+9)*pz
    dz = tl.load(CAM+10)*px + tl.load(CAM+11)*py + tl.load(CAM+12)*pz
    ox, oy, oz = tl.load(CAM+13), tl.load(CAM+14), tl.load(CAM+15)
    t = tl.full((BLOCK,), .002, tl.float32)
    hit = tl.full((BLOCK,), False, tl.int1)
    active = valid & (not BACKGROUND_ONLY)
    step = 0
    while (tl.sum(active.to(tl.int32), 0) > 0) & (step < STEPS):
        dist = field(ox+t*dx, oy+t*dy, oz+t*dz, BALL, ROD)
        # abs also handles an origin inside a scaffold solid: find the exit.
        new_hit = active & (tl.abs(dist) < .00015)
        hit = hit | new_hit
        active = active & ~new_hit
        t = tl.where(active, t + tl.maximum(tl.abs(dist)*.95, .000075), t)
        active = active & (t < FAR)
        step += 1
    x, y, z = ox+t*dx, oy+t*dy, oz+t*dz
    e = .0003
    nx = field(x+e,y,z,BALL,ROD)-field(x-e,y,z,BALL,ROD)
    ny = field(x,y+e,z,BALL,ROD)-field(x,y-e,z,BALL,ROD)
    nz = field(x,y,z+e,BALL,ROD)-field(x,y,z-e,BALL,ROD)
    nn = tl.rsqrt(nx*nx+ny*ny+nz*nz+1.e-12)
    diffuse = .55 + .45*tl.maximum(0., (nx*.36-ny*.80-nz*.48)*nn)
    qx, qy, qz = x-tl.floor(x+.5), y-tl.floor(y+.5), z-tl.floor(z+.5)
    sphere = tl.sqrt(qx*qx+qy*qy+qz*qz)-BALL < .0005
    # Directional hue map: independent of translation and scaffold spacing.
    lon = lib.atan2(dx, dz)
    lat = lib.asin(tl.minimum(1., tl.maximum(-1., -dy)))
    hue = lon/6.283185307179586+.5
    rr = tl.minimum(1., tl.maximum(0., tl.abs(hue*6.-3.)-1.))
    gg = tl.minimum(1., tl.maximum(0., 2.-tl.abs(hue*6.-2.)))
    bb = tl.minimum(1., tl.maximum(0., 2.-tl.abs(hue*6.-4.)))
    sat = .68 * (1.-.35*tl.abs(lat)/1.5707963267948966)
    val = .86 + .10*tl.cos(lat*2.)
    er, eg, eb = val*(1.-sat+sat*rr), val*(1.-sat+sat*gg), val*(1.-sat+sat*bb)
    if DIRECTION_RGB:
        # Continuous world-direction map, including at the poles. The base
        # RGB map distinguishes signed latitude; graticule is decorative.
        er, eg, eb = (dx+1.)*.5, (dy+1.)*.5, (dz+1.)*.5
    # Fine white angular grid matches the visual motif of Wonder Fig. 4.
    angle = 0.2617993877991494
    glon = tl.abs(lon/angle-tl.floor(lon/angle+.5))*angle*tl.cos(lat)
    glat = tl.abs(lat/angle-tl.floor(lat/angle+.5))*angle
    aa = .75/tl.load(CAM)
    grid = 1.-tl.minimum(1., tl.maximum(0., (tl.minimum(glon,glat)-aa)/(aa+1.e-7)))
    if DIRECTION_RGB:
        # Longitude is undefined at the poles: fade the angular overlay there.
        grid *= tl.minimum(1., tl.sqrt(dx*dx+dz*dz)/.1)
    er, eg, eb = er+(1.-er)*grid*.55, eg+(1.-eg)*grid*.55, eb+(1.-eb)*grid*.55
    cr = tl.where(sphere, .92, .26)*diffuse
    cg = tl.where(sphere, .28, .29)*diffuse
    cb = tl.where(sphere, .16, .31)*diffuse
    # Far attenuation reduces subpixel aliasing; world geometry remains fixed.
    alpha = tl.where(hit, tl.exp(-tl.maximum(t-2.,0.)/8.), 0.)
    cr, cg, cb = er*(1.-alpha)+cr*alpha, eg*(1.-alpha)+cg*alpha, eb*(1.-alpha)+cb*alpha
    tl.store(OUT+3*i, (255.*cr).to(tl.uint8), valid)
    tl.store(OUT+3*i+1, (255.*cg).to(tl.uint8), valid)
    tl.store(OUT+3*i+2, (255.*cb).to(tl.uint8), valid)
    tl.store(DEPTH+i, tl.where(hit,t,tl.where(active,-2.,-1.)),valid)


class Renderer:
    def __init__(self, width=832, height=480, aa=2, far=64., steps=4096, environment='direction_rgb'):
        self.width,self.height,self.aa,self.far,self.steps=width,height,aa,far,steps
        self.environment=environment
        self.rgb=torch.empty((height*aa,width*aa,3),device='cuda',dtype=torch.uint8)
        self.depth=torch.empty((height*aa,width*aa),device='cuda')

    def frame(self, c2w, K, spacing, radius=.05, rod=.008, background_only=False):
        # The initial camera lies at a cell center, for every spacing variant.
        # This phase is fixed for the whole trajectory, never recentered.
        # Reduce only the evaluation coordinate of the periodic renderer.
        # This is algebraically identical to the fixed infinite lattice, and
        # avoids losing precision in X-round(X). Keep actual memory addresses
        # unwrapped; this modulo is not a camera-following grid translation.
        center=np.remainder(c2w[:3,3].astype(np.float64)/spacing + .5, 1.)
        k=K.copy();k[:2,:]*=self.aa
        k[0,2]+=(self.aa-1)*.5;k[1,2]+=(self.aa-1)*.5
        packed=np.r_[k[0,0],k[1,1],k[0,2],k[1,2],c2w[:3,:3].ravel(),center].astype('float32')
        cam=torch.tensor(packed,device='cuda')
        w,h=self.width*self.aa,self.height*self.aa
        render_kernel[(tr.cdiv(w*h,128),)](cam,self.rgb,self.depth,w,h,radius,rod,self.far,self.steps,128,
                                         self.environment=='direction_rgb',background_only)
        rgb=self.rgb.cpu().numpy()
        if self.aa>1: rgb=cv2.resize(rgb,(self.width,self.height),interpolation=cv2.INTER_AREA)
        dep=self.depth[::self.aa,::self.aa].cpu().numpy()
        return rgb,dep


def make_trajectories(root, out, items=(4,10,15)):
    out.mkdir(parents=True,exist_ok=True)
    rows=[json.loads(x) for x in (root/'test/000000/index.jsonl').read_text().splitlines()]
    paths=[]
    for item in items:
        row=next(x for x in rows if x['source_item_index']==item)
        d=np.load(root/row['cameras']); t=d['timestamps_us'].astype(float)/1e6;t-=t[0]
        assert t[-1]>=5., (item,t[-1])
        target=np.arange(81)/16.;poses=d['c2w'].astype(float)
        result=np.repeat(np.eye(4)[None],81,axis=0)
        result[:,:3,:3]=Slerp(t,Rotation.from_matrix(poses[:,:3,:3]))(target).as_matrix()
        for j in range(3):result[:,j,3]=np.interp(target,t,poses[:,j,3])
        # One rigid gauge transform only; preserve metric scale.
        result=np.linalg.inv(result[0])[None]@result
        K=np.stack([np.interp(target,t,d['intrinsics'][:,i,j]) for i in range(3) for j in range(3)],axis=-1).reshape(81,3,3)
        h,w=d['image_size_hw'];scale=max(832/w,480/h)
        cropx=(w*scale-832)/2; cropy=(h*scale-480)/2
        K[:,:2,:]*=scale
        K[:,0,2]+=(scale-1)*.5-cropx;K[:,1,2]+=(scale-1)*.5-cropy
        name=f're10k_{item:02d}_{row["source_key"]}'
        path=out/(name+'.npz')
        np.savez(path,c2w=result.astype('float32'),K=K.astype('float32'),times_s=target)
        meta={'source_camera_file':str(root/row['cameras']),'source_metadata':str(root/row['metadata']),
              'source_image_hw':d['image_size_hw'].tolist(),'resize_scale':scale,'crop_xy':[cropx,cropy],
              'units':'official ViPE metric estimate, no translation scale fitting','sampling':'first 5 seconds, rotation Slerp, translation/K linear interpolation at 16fps',
              'translation_path_m':float(np.linalg.norm(np.diff(result[:,:3,3],axis=0),axis=1).sum())}
        path.with_suffix('.json').write_text(json.dumps(meta,indent=2));paths.append(path)
    K=np.repeat(np.array([[600,0,415.5],[0,600,239.5],[0,0,1]],dtype='float32')[None],81,axis=0)
    u=np.linspace(0,1,81)
    for name in ('translate_1m','rotate_60deg','closed_loop'):
        C=np.repeat(np.eye(4)[None],81,axis=0)
        if name=='translate_1m':C[:,0,3]=u
        elif name=='rotate_60deg':C[:,:3,:3]=Rotation.from_euler('y',60*u,degrees=True).as_matrix()
        else:
            C[:,0,3]=.4*np.sin(2*np.pi*u);C[:,2,3]=.3*(1-np.cos(2*np.pi*u))
            C[:,:3,:3]=Rotation.from_euler('y',20*np.sin(2*np.pi*u),degrees=True).as_matrix()
        path=out/(name+'.npz');np.savez(path,c2w=C.astype('float32'),K=K,times_s=u*5);paths.append(path)
    return paths


def writer(path,w=832,h=480):
    ffmpeg=shutil.which('ffmpeg') or '/opt/tiger/ss_bin/ffmpeg'
    return subprocess.Popen([ffmpeg,'-v','error','-y','-f','rawvideo','-pix_fmt','rgb24','-s',f'{w}x{h}',
                             '-r','16','-i','-','-an','-c:v','libx264','-crf','17','-preset','fast',
                             '-threads','4','-pix_fmt','yuv420p','-movflags','+faststart',str(path)],stdin=subprocess.PIPE)


def run(args):
    out=Path(args.output);out.mkdir(parents=True,exist_ok=True)
    paths=make_trajectories(Path(args.annotations),out/'trajectories',args.scene_items)
    renderer=Renderer(aa=args.aa,far=args.far,steps=args.steps,environment=args.environment)
    tasks=[]
    if args.mode=='preview':
        for spacing in (.25,.5,1.,2.):
            rgb,dep=renderer.frame(np.eye(4),np.array([[600,0,415.5],[0,600,239.5],[0,0,1.]]),spacing,args.radius)
            cv2.imwrite(str(out/f'preview_{spacing}.png'),cv2.cvtColor(rgb,cv2.COLOR_RGB2BGR))
            print(spacing,{'hit':float((dep>0).mean()),'unresolved':float((dep==-2).mean())},flush=True)
        return
    for path in paths[:len(args.scene_items)]:
        for spacing in (.25,.5,1.,2.):tasks.append((path,spacing,args.radius))
    for radius in args.radius_extras:tasks.append((paths[0],.5,radius))
    if not args.no_synthetic_videos:
        for path in paths[len(args.scene_items):]:tasks.append((path,.5,args.radius))
    (out/'videos').mkdir(exist_ok=True);(out/'stills').mkdir(exist_ok=True)
    summaries=[]
    for path,spacing,radius in tasks:
        name=f'{path.stem}_d{spacing:g}_r{radius:g}';video=out/'videos'/(name+'.mp4')
        d=np.load(path);stats=[];start=time.perf_counter();proc=writer(video)
        try:
            for t in range(81):
                rgb,dep=renderer.frame(d['c2w'][t],d['K'][t],spacing,radius)
                proc.stdin.write(rgb.tobytes())
                if t in (0,20,40,60,80):cv2.imwrite(str(out/'stills'/f'{name}_{t:03d}.png'),cv2.cvtColor(rgb,cv2.COLOR_RGB2BGR))
                stats.append({'hit':float((dep>0).mean()),'unresolved':float((dep==-2).mean()),
                              'beyond_4_cells':float((dep>4).mean()),'beyond_8_cells':float((dep>8).mean()),
                              'near_half_cell':float(((dep>0)&(dep<.5)).mean())})
        finally:proc.stdin.close()
        if proc.wait()!=0:raise RuntimeError('ffmpeg failed')
        summary={'name':name,'video':str(video),'trajectory':str(path),'spacing_m':spacing,'sphere_radius_over_spacing':radius,
                 'rod_radius_over_spacing':.008,'seconds':time.perf_counter()-start,
                 'mean_hit':float(np.mean([s['hit'] for s in stats])),
                 'max_unresolved':max(s['unresolved'] for s in stats),
                 'max_near_half_cell':max(s['near_half_cell'] for s in stats),
                 'mean_beyond_4_cells':float(np.mean([s['beyond_4_cells'] for s in stats])),
                 'mean_beyond_8_cells':float(np.mean([s['beyond_8_cells'] for s in stats]))}
        summaries.append(summary)
        (out/'metrics.json').write_text(json.dumps(summaries,indent=2))
        print(json.dumps(summary),flush=True)
    manifest={'worker':'1094200','shape':[81,480,832],'fps':16,'aa':args.aa,'far_cells':args.far,'max_steps':args.steps,
              'geometry':'infinite periodic sphere/cylinder union, finite ray distance','origin':'(-Delta/2,-Delta/2,-Delta/2), fixed per trajectory',
              'renderer':'custom CUDA Triton sphere tracing; not official Wonder OpenGL code','device':torch.cuda.get_device_name(),
              'environment_mode':args.environment,'base_radius_ratio':args.radius,
              'environment':('E(d)=(d+1)/2; 15 degree white graticule with polar fade' if args.environment=='direction_rgb'
                             else 'world longitude hue + latitude modulation + 15 degree white graticule'),
              'distance_fog':'exp(-max(ray_distance_cells-2,0)/8), blends into directional environment','variants':summaries}
    (out/'manifest.json').write_text(json.dumps(manifest,indent=2))


if __name__=='__main__':
    p=argparse.ArgumentParser(description=__doc__)
    p.add_argument('--output',required=True)
    p.add_argument('--annotations',default='/mnt/hdfs/data_ssd/tianqi/re10k_annotations/vipe_dav3_v1')
    p.add_argument('--scene-items',type=int,nargs='+',default=[4,10,15])
    p.add_argument('--radius',type=float,default=.05)
    p.add_argument('--radius-extras',type=float,nargs='*',default=[])
    p.add_argument('--environment',choices=['direction_rgb','legacy_hsv'],default='direction_rgb')
    p.add_argument('--no-synthetic-videos',action='store_true')
    p.add_argument('--aa',type=int,default=2);p.add_argument('--far',type=float,default=64.)
    p.add_argument('--steps',type=int,default=4096);p.add_argument('--mode',choices=['preview','sweep'],default='sweep')
    run(p.parse_args())