def areDoorsInRoom2021(level):
level_doorfix = level.copy()
# for each room in level, check each door;
for room in level_doorfix['rooms']:
# inDatabase Check:
for o in room['objList']:
if o['modelId'] in sk_to_ali or o['modelId'] in suncg:
o['inDatabase'] = True
else:
o['inDatabase'] = False
if not os.path.exists('room/{}/{}f.obj'.format(room['origin'], room['modelId'])):
continue
try:
room_meta = p2d('.', 'room/{}/{}f.obj'.format(room['origin'], room['modelId']))
room_polygon = Polygon(room_meta[:, 0:2]) # requires python library 'shapely'
except Exception as e:
print(e)
continue
for r in level['rooms']:
for obj in r['objList']:
if obj is None:
continue
if 'coarseSemantic' not in obj:
continue
if obj['coarseSemantic'] not in ['door', 'Door']:
continue
block = windoorblock_f(obj)
block_polygon = Polygon(block['windoorbb']).buffer(.03)
# for this time, we do not duplicate doors, instead we add roomIds to the obj.
if room_polygon.intersects(block_polygon):
if 'roomIds' not in obj:
obj['roomIds'] = []
obj['roomIds'].append(room['roomId'])
return level_doorfix
def fa_reshuffle(rj):
pend_obj_list = []
final_obj_list = []
bbindex = []
ol = rj['objList']
print("Total Number of objects: ", len(ol))
for o in ol:
if o is None:
continue
if o['modelId'] not in obj_semantic:
final_obj_list.append(o)
continue
if 'coarseSemantic' in o:
if o['coarseSemantic'] in BANNED:
final_obj_list.append(o)
continue
bbindex.append(name_to_ls[o['modelId']])
pend_obj_list.append(o)
room_meta = p2d(
'.', '/suncg/room/{}/{}f.obj'.format(rj['origin'], rj['modelId']))
room_polygon = Polygon(room_meta[:, 0:2])
room_shape = torch.from_numpy(room_meta[:, 0:2]).float()
room_shape_norm = torch.from_numpy(room_meta).float()
for o in pend_obj_list:
disturbance(o, 0.5, room_polygon)
for o in pend_obj_list:
o['rotate'][0] = 0.0
o['rotate'][1] = o['orient']
o['rotate'][2] = 0.0
return rj
def floorplanOrthes():
pt.cameraType = 'orthographic'
pt.SAVECONFIG = False
pt.REMOVELAMP = True
floorplanlist = os.listdir('./dataset/alilevel_door2021/')
# for floorplanfile in floorplanlist:
for floorplanfile in [
'e8b0a6bf-58a2-49de-b9ea-231995fc9e3b.json',
'317d64ff-b96e-4743-88f6-2b5b27551a7c.json'
]:
if '.json' not in floorplanfile:
continue
with open(f'./dataset/alilevel_door2021/{floorplanfile}') as f:
scenejson = json.load(f)
# if os.path.exists(f"./dataset/alilevel_door2021_orth/{scenejson['origin']}.png"):
# continue
points = []
for room in scenejson['rooms']:
try:
floorMeta = p2d(
'.',
'/dataset/room/{}/{}f.obj'.format(room['origin'],
room['modelId']))
points += floorMeta[:, 0:2].tolist()
wallMeta = sk.getMeshVertices(
'/dataset/room/{}/{}w.obj'.format(room['origin'],
room['modelId']))
points += wallMeta[:, [0, 2]].tolist()
except:
continue
v = np.array(points)
l = np.min(v[:, 0])
r = np.max(v[:, 0])
u = np.min(v[:, 1])
d = np.max(v[:, 1])
# orthViewLen = max((r - l), (d - u)) + 0.45
orthViewLen = (r - l) + 0.45
scenejson["PerspectiveCamera"] = {}
scenejson["PerspectiveCamera"]['origin'] = [(r + l) / 2, 50,
(d + u) / 2]
scenejson["PerspectiveCamera"]['target'] = [(r + l) / 2, 0,
(d + u) / 2]
scenejson["PerspectiveCamera"]['up'] = [0, 0, 1]
scenejson["OrthCamera"] = {}
scenejson["OrthCamera"]['x'] = orthViewLen / 2
scenejson["OrthCamera"]['y'] = orthViewLen / 2
scenejson["canvas"] = {}
scenejson['canvas']['width'] = int((r - l) * 100)
scenejson['canvas']['height'] = int((d - u) * 100)
print(f'Rendering {floorplanfile} ...')
try:
pt.pathTracing(
scenejson, 64,
f"./dataset/alilevel_door2021_orth/{scenejson['origin']}.png")
except Exception as e:
print(e)
continue
# swap the cameraType back to perspective cameras.
pt.cameraType = 'perspective'
def mageAddAuto():
rj = request.json
if 'auxiliaryDomObj' in rj:
room_meta = rj['auxiliaryDomObj']['room_meta']
else:
room_meta = p2d('.', f'/dataset/room/{rj["origin"]}/{rj["modelId"]}f.obj')
samples = random_points_within(Polygon(room_meta), 1000)
print(samples[0])
return json.dumps(samples)
def alilevel2room():
AREA = {}
levelnames = os.listdir('./dataset/alilevel_windoorFix')
for levelname in levelnames:
with open(f'./dataset/alilevel_windoorFix/{levelname}') as f:
level = json.load(f)
for room in level['rooms']:
if len(room['roomTypes']) == 0:
roomtype = 'Unknown'
else:
roomtype = room['roomTypes'][0]
newlevel = level.copy()
newlevel['rooms'] = [room]
if (len(room['objList']) != 0):
newlevel['bbox'] = room['bbox']
newlevel['rooms'][0]['roomId'] = 0
try:
newlevel['rooms'][0]['area'] = Polygon(
p2d(f'./dataset/room/{newlevel["origin"]}',
f'{newlevel["rooms"][0]["modelId"]}f.obj')).area
except Exception as e:
newlevel['rooms'][0]['area'] = 0.0
print(e)
for o in newlevel['rooms'][0]['objList']:
o['roomId'] = 0
if o['modelId'] in sk_to_ali or o['modelId'] in suncg:
o['inDatabase'] = True
else:
o['inDatabase'] = False
if not os.path.exists(f'./dataset/alilevel_inRooms/{roomtype}'):
os.makedirs(f'./dataset/alilevel_inRooms/{roomtype}')
if roomtype not in AREA:
AREA[roomtype] = {}
samefileindex = 0
finalfilename = f'./dataset/alilevel_inRooms/{roomtype}/{roomtype}-{newlevel["origin"]}-{samefileindex}.json'
while os.path.exists(finalfilename):
samefileindex += 1
finalfilename = f'./dataset/alilevel_inRooms/{roomtype}/{roomtype}-{newlevel["origin"]}-{samefileindex}.json'
with open(finalfilename, 'w') as f:
json.dump(newlevel, f)
AREA[roomtype][
f'{roomtype}-{newlevel["origin"]}-{samefileindex}'] = newlevel[
'rooms'][0]['area']
with open('./dataset/AREA.json', 'w') as f:
json.dump(AREA, f)
def balancing(h, room, theta):
"""
'h' is a generated probe view.
"""
h['direction'] /= np.linalg.norm(h['direction'])
floorMeta = p2d(
'.', '/dataset/room/{}/{}f.obj'.format(room['origin'],
room['modelId']))
onePlanePointList = []
for obj in room['objList']:
if not isObjectInSight(obj, h['probe'], h['direction'], floorMeta,
theta, room['objList'], False):
continue
probeTot = np.array(obj['translate']) - h['probe']
cosToDirection = np.dot(probeTot,
h['direction']) / np.linalg.norm(probeTot)
DIS = 1 / cosToDirection
DRC = probeTot / np.linalg.norm(probeTot)
onePlanePointList.append(h['probe'] + DIS * DRC)
centroid = sum(onePlanePointList) / len(onePlanePointList)
newDirection = centroid - h['probe']
newDirection /= np.linalg.norm(newDirection, ord=2)
return newDirection
def autoViewsRodrigues(room, scene):
# change the fov/2 to Radian.
theta = (np.pi * scene['PerspectiveCamera']['fov'] / 180) / 2
# the the floor meta.
floorMeta = p2d(
'.', '/dataset/room/{}/{}f.obj'.format(room['origin'],
room['modelId']))
floorPoly = Polygon(floorMeta[:, 0:2])
# the height of the wall.
H = sk.getWallHeight(
f"./dataset/room/{room['origin']}/{room['modelId']}w.obj")
pcams = []
for wallIndex in range(floorMeta.shape[0]):
pcam = {}
wallIndexNext = (wallIndex + 1) % floorMeta.shape[0]
middlePoint = (floorMeta[wallIndex][0:2] +
floorMeta[wallIndexNext][0:2]) / 2
middlePoint += floorMeta[wallIndex][2:4] * 0.005
origin = middlePoint.tolist()
origin.insert(1, H / 2)
direction = floorMeta[wallIndex][2:4].tolist()
direction.insert(1, 0.)
origin = np.array(origin)
direction = np.array(direction)
pcam['theta'] = theta
pcam['roomId'] = room['roomId']
# pcam['viewLength'] = np.linalg.norm(middlePoint - p, ord=2)
pcam['probe'] = origin
pcam['wallIndex'] = wallIndex
pcam['direction'] = groundShifting(origin, floorMeta, floorPoly,
direction, theta, H)
pcam['type'] = 'againstMidWall'
pcams.append(pcam)
return pcams
def sceneSynthesis(rj):
print(rj['origin'])
start_time = time.time()
pend_obj_list = []
bbindex = []
blocks = []
random.shuffle(rj['objList'])
# bounding boxes of given furniture objects;
boundingboxes = []
max_points = []
min_points = []
# identifying objects to arrange;
for o in rj['objList']:
if o is None:
print('this is a None object; ')
continue
if 'coarseSemantic' not in o:
print('a given object does not have coarseSemantic; ')
continue
if o['coarseSemantic'] in BANNED:
print('a given object is not a furniture;' )
continue
if o['coarseSemantic'] == 'door' or o['coarseSemantic'] == 'window':
blocks.append(windoorblock_f(o))
continue
# if o['modelId'] not in obj_semantic:
# print(f'a given object {o["modelId"]} is not a furniture;' )
# continue
# bbindex.append(name_to_ls[o['modelId']])
try:
boundingboxes.append(load_boundingbox(o['modelId']))
except Exception as e:
continue
aabb = load_AABB(o['modelId'])
max_points.append(aabb['max'])
min_points.append(aabb['min'])
o['childnum'] = {}
o['myparent'] = None
pend_obj_list.append(o)
# load priors;
csrrelation = torch.zeros((len(pend_obj_list), len(pend_obj_list)), dtype=torch.float)
for center in pend_obj_list:
for obj in pend_obj_list:
preload_prior(center['modelId'], obj['modelId'])
for centerid in range(len(pend_obj_list)):
center = pend_obj_list[centerid]
for objid in range(len(pend_obj_list)):
if objid == centerid:
continue
obj = pend_obj_list[objid]
# if the obj has a parent, we have to continue;
# because if multiple parents exist, two parent may share a same child while another child has no parent;
if obj['myparent'] is not None:
continue
pid = "{}-{}".format(center['modelId'], obj['modelId'])
if pid in priors['pos']:
if obj['modelId'] not in center['childnum']:
center['childnum'][obj['modelId']] = 0
if center['childnum'][obj['modelId']] >= priors['chainlength'][pid]:
continue
csrrelation[centerid, objid] = 1.
obj['myparent'] = center
center['childnum'][obj['modelId']] += 1
# partition coherent groups;
pend_groups = connected_component(np.arange(len(pend_obj_list)), csrrelation)
cgs = []
for pend_group in pend_groups:
cg = {}
cg['objList'] = [pend_obj_list[i] for i in pend_group]
cg['csrrelation'] = csrrelation[pend_group][:, pend_group]
cg['translate'] = [0.0, 0.0, 0.0]
cg['orient'] = 0.0
# determine layouts of each group;
heuristic(cg)
# the following code is for chain pattern;
# if only one object exists in a cg && the object follows chain layout,
# e.g., kitchen cabinet, shelving, etc;
if len(cg['objList']) == 1 and cg['objList'][0]['coarseSemantic'] in NaiveChainList:
cg['chain'] = cg['objList'][0]['coarseSemantic']
else:
cg['chain'] = 'n'
if cg['objList'][cg['leaderID']]['modelId'] in ['781'] and cg['objList'][cg['leaderID']]['translate'][1] == 0:
cg['objList'][cg['leaderID']]['translate'][1] = 1.04
cgs.append(cg)
# load and process room shapes;
room_meta = p2d('.', '/dataset/room/{}/{}f.obj'.format(rj['origin'], rj['modelId']))
room_polygon = Polygon(room_meta[:, 0:2]) # requires python library 'shapely'
translate = torch.zeros((len(pend_obj_list), 3)).float()
orient = torch.zeros((len(pend_obj_list))).float()
scale = torch.zeros((len(pend_obj_list), 3)).float()
for i in range(len(pend_obj_list)):
translate[i][0] = pend_obj_list[i]['translate'][0]
translate[i][1] = pend_obj_list[i]['translate'][1]
translate[i][2] = pend_obj_list[i]['translate'][2]
orient[i] = pend_obj_list[i]['orient']
scale[i][0] = pend_obj_list[i]['scale'][0]
scale[i][1] = pend_obj_list[i]['scale'][1]
scale[i][2] = pend_obj_list[i]['scale'][2]
# bb = four_points_xz[bbindex].float()
# max_points = max_bb[bbindex].float()
# min_points = min_bb[bbindex].float()
bb = torch.tensor(boundingboxes).float()
max_points = torch.tensor(max_points).float()
min_points = torch.tensor(min_points).float()
for i in range(len(pend_obj_list)):
bb[i] = rotate_bb_local_para(bb[i], orient[i], scale[i][[0, 2]])
max_points[i] = transform_a_point(max_points[i], translate[i], orient[i], scale[i])
min_points[i] = transform_a_point(min_points[i], translate[i], orient[i], scale[i])
bb_tran = translate.reshape(len(pend_obj_list), 1, 3)[:, :, [0, 2]] + bb # note that bbs are around (0,0,0) after heuristic(cg)
# calculate bounding box of coherent groups;
for gid in range(len(pend_groups)):
pend_group = pend_groups[gid]
cg = cgs[gid]
points = bb_tran[pend_group].reshape(-1, 2)
max_points_of_cg = max_points[pend_group]
min_points_of_cg = min_points[pend_group]
maxp = torch.max(points, dim=0)[0]
minp = torch.min(points, dim=0)[0]
cg['bb'] = torch.zeros((4, 2), dtype=torch.float)
cg['bb'][0] = maxp
cg['bb'][1][0] = minp[0]
cg['bb'][1][1] = maxp[1]
cg['bb'][2] = minp
cg['bb'][3][0] = maxp[0]
cg['bb'][3][1] = minp[1]
cg['height'] = torch.max(max_points_of_cg, dim=0)[0][1].item()
cg['ground'] = torch.min(min_points_of_cg, dim=0)[0][1].item()
# generate
attempt_heuristic(cgs, room_meta, blocks)
for cg in cgs:
# the following code is reserving for lifting of each coherent group;
cg['translate'][0] += np.sin(cg['orient']) * 0.08
cg['translate'][2] += np.cos(cg['orient']) * 0.08
# if cg['objList'][cg['leaderID']]['coarseSemantic'] in ['sink', 'dressing_table', 'picture_frame', 'television', 'mirror', 'clock', 'dining_table']:
# cg['translate'][0] += np.sin(cg['orient']) * 0.08
# cg['translate'][2] += np.cos(cg['orient']) * 0.08
for o in cg['objList']:
o['translate'][0], o['translate'][2] = rotate([0, 0], [o['translate'][0], o['translate'][2]], cg['orient'])
o['orient'] += cg['orient']
o['rotate'][0] = 0.0
o['rotate'][1] = o['orient']
o['rotate'][2] = 0.0
o['translate'][0] += cg['translate'][0]
o['translate'][1] += cg['translate'][1]
o['translate'][2] += cg['translate'][2]
# log coherent groups;
for i in range(len(pend_obj_list)):
o = pend_obj_list[i]
if 'coarseSemantic' not in o:
break
print(o['modelId'], o['coarseSemantic'])
for j in range(len(pend_obj_list)):
if csrrelation[i][j] == 1.0:
print("--->>>", pend_obj_list[j]['modelId'], pend_obj_list[j]['coarseSemantic'])
print("\r\n--- %s secondes ---" % (time.time() - start_time))
return rj
def priors_of_roomShape():
rj = request.json
existingCatList = []
for obj in rj['objList']:
if obj is None:
continue
if 'modelId' not in obj:
continue
if obj['modelId'] not in objCatList:
continue
if len(objCatList[obj['modelId']]) == 0:
continue
if objCatList[obj['modelId']][0] not in existingCatList:
existingCatList.append(objCatList[obj['modelId']][0])
existingPendingCatList = existingCatList.copy()
res = {'object': [], 'prior': [], 'index': [], 'coarseSemantic': {}, 'catMask': []}
if 'auxiliaryDomObj' in rj:
if 'heyuindex' in rj['auxiliaryDomObj']:
res['heyuindex'] = rj['auxiliaryDomObj']['heyuindex']
for objname in rj['auxiliaryDomObj']['object']:
if objCatList[objname][0] not in existingPendingCatList:
existingPendingCatList.append(objCatList[objname][0])
# print(existingCatList)
# print(existingPendingCatList)
# load and process room shapes;
room_meta = p2d('.', f'/dataset/room/{rj["origin"]}/{rj["modelId"]}f.obj')
room_meta = room_meta[:, 0:2]
wallSecIndices = np.arange(1, len(room_meta)).tolist() + [0]
res['room_meta'] = room_meta.tolist()
rv = room_meta[:] - room_meta[wallSecIndices]
normals = rv[:, [1,0]]
normals[:, 1] = -normals[:, 1]
res['room_orient'] = np.arctan2(normals[:, 0], normals[:, 1]).tolist()
res['room_oriNormal'] = normals.tolist()
# room_polygon = Polygon(room_meta[:, 0:2]) # requires python library 'shapely'
# currently, we hack few available coherent groups...
roomTypeSuggestedList = []
categoryList = []
for rt in rj['roomTypes']:
if 'heyuindex' not in res:
res['heyuindex'] = np.random.randint(len(roomTypeDemo[rt]))
categoryList += roomTypeDemo[rt][res['heyuindex']]
break
for cat in categoryList:
if cat in existingCatList:
continue
roomTypeSuggestedList.append(random.choice(objListCat[cat]))
if 'auxiliaryDomObj' not in rj: # if this is the first time calling this function for the pending room...
res['object'] = roomTypeSuggestedList
else:
# re-fuel the pending object list;
for newobjname in roomTypeSuggestedList:
if objCatList[newobjname][0] in existingPendingCatList:
continue
rj['auxiliaryDomObj']['object'].append(newobjname)
res['object'] = rj['auxiliaryDomObj']['object'].copy()
for objname in rj['auxiliaryDomObj']['object']:
# if the specific instance is already in the room;
if objCatList[objname][0] in existingCatList:
res['object'].remove(objname)
if len(res['object']) == 0: # if a recommendation list is full:
pass
# load wall priors;
for obj in res['object']:
with open(f'./latentspace/wdot-4/{obj}.json') as f:
wallpri = json.load(f)
res['prior'] += wallpri
res['index'] += np.full(len(wallpri), obj).tolist()
res['catMask'] += np.full(len(wallpri), categoryCodec[getobjCat(obj)]).tolist()
for newobjname in res['object']:
res['coarseSemantic'][newobjname] = getobjCat(newobjname)
return json.dumps(res)
def fa_layout_pro(rj):
pend_obj_list = []
final_obj_list = []
bbindex = []
ol = rj['objList']
print("Total Number of objects: ", len(ol))
for o in ol:
if o is None:
continue
if o['modelId'] not in obj_semantic:
final_obj_list.append(o)
continue
if 'coarseSemantic' in o:
if o['coarseSemantic'] in BANNED:
final_obj_list.append(o)
continue
bbindex.append(name_to_ls[o['modelId']])
pend_obj_list.append(o)
diag_indices_ = (torch.arange(len(pend_obj_list)),
torch.arange(len(pend_obj_list)))
csrrelation = csrmatrix[bbindex][:, bbindex]
yrelation = ymatrix[bbindex][:, bbindex]
wallrelation = wallvector[bbindex]
cornerrelation = cornervector[bbindex]
csrrelation[diag_indices_] = 0.0
SSIZE = 1000
rng = np.random.default_rng()
for centerid in range(len(pend_obj_list)):
center = pend_obj_list[centerid]
for objid in range(len(pend_obj_list)):
obj = pend_obj_list[objid]
priorid = "{}-{}".format(center['modelId'], obj['modelId'])
if priorid not in priors['pos']:
with open(PRIORS.format(center['modelId'])) as f:
if csrrelation[centerid, objid] == 0.0:
theprior = np.zeros((SSIZE, 4), dtype=np.float)
else:
theprior = np.array(json.load(f)[obj['modelId']],
dtype=np.float)
if len(theprior) == 0:
theprior = np.zeros((SSIZE, 4), dtype=np.float)
while len(theprior) < SSIZE:
theprior = np.vstack((theprior, theprior))
rng.shuffle(theprior)
priors['pos'][priorid] = theprior[:, 0:3]
priors['ori'][priorid] = theprior[:, 3].flatten()
SSIZE = np.min((len(priors['pos'][priorid]), SSIZE))
priors['pos'][priorid] = torch.from_numpy(
priors['pos'][priorid]).float()
priors['ori'][priorid] = torch.from_numpy(
priors['ori'][priorid]).float()
else:
SSIZE = np.min((len(priors['pos'][priorid]), SSIZE))
pos_priors = torch.zeros(len(pend_obj_list), len(pend_obj_list), SSIZE, 3)
ori_priors = torch.zeros(len(pend_obj_list), len(pend_obj_list), SSIZE)
for centerid in range(len(pend_obj_list)):
center = pend_obj_list[centerid]
for objid in range(len(pend_obj_list)):
if objid == centerid:
continue
obj = pend_obj_list[objid]
priorid = "{}-{}".format(center['modelId'], obj['modelId'])
pos_priors[centerid, objid] = rotate_pos_prior(
priors['pos'][priorid][0:SSIZE],
torch.tensor(center['orient'], dtype=torch.float))
ori_priors[centerid, objid] = priors['ori'][priorid][0:SSIZE]
# making sure that angles are between (-pi, pi)
ori_priors[ori_priors > np.pi] -= 2 * np.pi
ori_priors[ori_priors < -np.pi] += 2 * np.pi
room_meta = p2d(
'.', '/suncg/room/{}/{}f.obj'.format(rj['origin'], rj['modelId']))
room_polygon = Polygon(room_meta[:, 0:2])
room_shape = torch.from_numpy(room_meta[:, 0:2]).float()
room_shape_norm = torch.from_numpy(room_meta).float()
translate = torch.zeros((len(pend_obj_list), 2)).float()
orient = torch.zeros((len(pend_obj_list))).float()
scale = torch.zeros((len(pend_obj_list), 3)).float()
# for o in pend_obj_list:
# disturbance(o, 0.5, room_polygon)
for i in range(len(pend_obj_list)):
translate[i][0] = pend_obj_list[i]['translate'][0]
translate[i][1] = pend_obj_list[i]['translate'][2]
orient[i] = pend_obj_list[i]['orient']
scale[i][0] = pend_obj_list[i]['scale'][0]
scale[i][1] = pend_obj_list[i]['scale'][1]
scale[i][2] = pend_obj_list[i]['scale'][2]
bb = four_points_xz[bbindex].float()
for i in range(len(pend_obj_list)):
bb[i] = rotate_bb_local_para(bb[i], orient[i], scale[i][[0, 2]])
translate.requires_grad_()
orient.requires_grad_()
iteration = 0
# loss = distribution_loss(translate, pos_priors[:, :, :, [0, 2]], csrrelation)
start_time = time.time()
loss = distribution_loss_orient(translate, orient, pos_priors[:, :, :,
[0, 2]],
ori_priors, csrrelation)
c_loss = collision_loss(
translate.reshape(len(pend_obj_list), 1, 2) + bb, room_shape,
yrelation * (1 - csrrelation), wallrelation, cornerrelation)
loss += c_loss
while loss.item() > 0.0 and iteration < MAX_ITERATION:
print("Start iteration {}...".format(iteration))
loss.backward()
# translate.data = translate.data - (1.0 / (1 + torch.sum(csrrelation, dim=1))).reshape(len(pend_obj_list), 1) * translate.grad * 0.05
translate.data = translate.data - translate.grad * 0.05
translate.grad = None
# if orient.grad is not None:
# orient.data = orient.data - orient.grad * 0.01
# orient.data[orient.data > np.pi] -= 2 * np.pi
# orient.data[orient.data < -np.pi] += 2 * np.pi
# orient.grad = None
# loss = distribution_loss(translate, pos_priors[:, :, :, [0, 2]], csrrelation)
loss = distribution_loss_orient(translate, orient, pos_priors[:, :, :,
[0, 2]],
ori_priors, csrrelation)
c_loss = collision_loss(
translate.reshape(len(pend_obj_list), 1, 2) + bb, room_shape,
yrelation * (1 - csrrelation), wallrelation, cornerrelation)
loss += c_loss
iteration += 1
print("--- %s seconds ---" % (time.time() - start_time))
for i in range(len(pend_obj_list)):
o = pend_obj_list[i]
if 'coarseSemantic' not in o:
break
print(o['modelId'], o['coarseSemantic'])
for j in range(len(pend_obj_list)):
if csrrelation[i][j] == 1.0:
print("--->>>", pend_obj_list[j]['modelId'],
pend_obj_list[j]['coarseSemantic'])
print(csrrelation)
for i in range(len(pend_obj_list)):
o = pend_obj_list[i]
o['translate'][0] = translate[i][0].item()
o['translate'][2] = translate[i][1].item()
o['rotate'][0] = 0.0
o['rotate'][1] = orient[i].item()
o['rotate'][2] = 0.0
o['orient'] = orient[i].item()
return rj
def hamilton(scene):
involvedRoomIds = []
views = []
# load existing views.
for fn in os.listdir(f'./latentspace/autoview/{scene["origin"]}'):
if '.json' not in fn:
continue
with open(f'./latentspace/autoview/{scene["origin"]}/{fn}') as f:
views.append(json.load(f))
for view in views:
view['isVisited'] = False
if view['roomId'] not in involvedRoomIds:
involvedRoomIds.append(view['roomId'])
print(involvedRoomIds)
res = []
# deciding connections of a floorplan.
G = nx.Graph()
for room in scene['rooms']:
room['isVisited'] = False
floorMeta = p2d(
'.', '/dataset/room/{}/{}f.obj'.format(room['origin'],
room['modelId']))
try:
H = sk.getWallHeight(
f"./dataset/room/{room['origin']}/{room['modelId']}w.obj")
except:
continue
for door in room['objList']:
if 'coarseSemantic' not in door:
continue
if door['coarseSemantic'] not in ['Door', 'door']:
continue
if len(door['roomIds']) < 2:
continue
# if door['roomIds'][0] not in involvedRoomIds and door['roomIds'][1] not in involvedRoomIds:
# continue
x = (door['bbox']['min'][0] + door['bbox']['max'][0]) / 2
z = (door['bbox']['min'][2] + door['bbox']['max'][2]) / 2
DIS = np.Inf
for wallIndex in range(floorMeta.shape[0]):
wallIndexNext = (wallIndex + 1) % floorMeta.shape[0]
dis = sk.pointToLineDistance(np.array([x, z]),
floorMeta[wallIndex, 0:2],
floorMeta[wallIndexNext, 0:2])
if dis < DIS:
DIS = dis
direction = np.array(
[floorMeta[wallIndex, 2], 0, floorMeta[wallIndex, 3]])
translate = np.array([x, H / 2, z])
G.add_edge(door['roomIds'][0],
door['roomIds'][1],
translate=translate,
direction=direction,
directionToRoom=room['roomId'])
pre = nx.dfs_predecessors(G)
suc = nx.dfs_successors(G)
print(pre, suc)
# decide the s and t which are the start point and end point respectively.
# ndproom = list(nx.dfs_successors(G).keys())[0]
# ndproom = views[0]['roomId']
ndproom = involvedRoomIds[0]
roomOrder = []
while ndproom != -1:
roomOrder.append(ndproom)
scene['rooms'][ndproom]['isVisited'] = True
ndproom = hamiltonNextRoom(ndproom, pre, suc, scene)
for room in scene['rooms']:
room['isVisited'] = False
print(roomOrder)
def subPath(s):
if s == len(roomOrder) - 1:
return (True, s)
state = False
start = roomOrder[s]
s += 1
while s < len(roomOrder) and roomOrder[s] != start:
if roomOrder[s] in involvedRoomIds and not scene['rooms'][
roomOrder[s]]['isVisited']:
state = True
s += 1
return (state, s)
i = 0
while i < len(roomOrder):
state, s = subPath(i)
if not state:
roomOrder = roomOrder[0:i + 1] + roomOrder[s + 1:]
i -= 1
else:
scene['rooms'][roomOrder[i]]['isVisited'] = True
i += 1
print(roomOrder)
ndproom = roomOrder[0]
for view in views:
if view['roomId'] == ndproom:
ndpNext = view
# perform the algorithm of Angluin and Valiant.
for i in range(1, len(roomOrder) + 1):
while ndpNext is not None:
ndp = ndpNext
res.append(ndp)
ndp['isVisited'] = True
ndpNext = hamiltonNext(ndp, views, scene)
if i == len(roomOrder):
break
lastndproom = roomOrder[i - 1]
ndproom = roomOrder[i]
edge = G[lastndproom][ndproom]
# if edge['direction'].dot(edge['translate'] - ndp['probe']) < 0:
if edge['directionToRoom'] != ndproom:
edge['direction'] = -edge['direction']
ndpNext = {
'roomId': ndproom,
'probe': edge['translate'],
'origin': edge['translate'].tolist(),
'target': (edge['translate'] + edge['direction']).tolist(),
'direction': edge['direction'].tolist()
}
with open(f'./latentspace/autoview/{scene["origin"]}/path', 'w') as f:
json.dump(res, f, default=sk.jsonDumpsDefault)
return res
def autoViewOnePointPerspective(room, scene):
"""
This function tries generate all potential views w.r.t the One-Point Perspective Rule (OPP Rule).
Note that several variants exist w.r.t different rules.
"""
theta = (np.pi * scene['PerspectiveCamera']['fov'] / 180) / 2
floorMeta = p2d(
'.', '/dataset/room/{}/{}f.obj'.format(room['origin'],
room['modelId']))
floorPoly = Polygon(floorMeta[:, 0:2])
H = sk.getWallHeight(
f"./dataset/room/{room['origin']}/{room['modelId']}w.obj")
# MAXDIAMETER = sk.roomDiameter(floorMeta)
# find the anchor point and the anchor wall.
hypotheses = []
hypotheses += autoViewsRodrigues(room, scene)
hypotheses += autoViewTwoWallPerspective(room, scene)
for wallIndex in range(floorMeta.shape[0]):
# first get the meta from the target wall.
wallIndexNext = (wallIndex + 1) % floorMeta.shape[0]
middlePoint = (floorMeta[wallIndex][0:2] +
floorMeta[wallIndexNext][0:2]) / 2
normal = floorMeta[wallIndex][2:4]
normal3D = np.array([normal[0], 0, normal[1]])
# construct the probe lineString.
p1 = middlePoint
p2 = middlePoint + normal
# detect wall decorations including windows.
targetWallNumWindows = 0
for r in scene['rooms']:
for obj in r['objList']:
if isWindowOnWall(obj, floorMeta[wallIndex][0:2],
floorMeta[wallIndexNext][0:2]):
targetWallNumWindows += 1
targetWallWindoorArea = 0.
for r in scene['rooms']:
for obj in r['objList']:
targetWallWindoorArea += calWindoorArea(
obj, floorMeta[wallIndex][0:2],
floorMeta[wallIndexNext][0:2])
for wallJndex in range(floorMeta.shape[0]):
if wallJndex == wallIndex:
continue
if np.dot(floorMeta[wallIndex][2:4],
floorMeta[wallJndex][2:4]) >= 0:
continue
p3 = floorMeta[wallJndex][0:2]
p4 = floorMeta[(wallJndex + 1) % floorMeta.shape[0]][0:2]
# generate the probe point.
p = twoInfLineIntersection(p1, p2, p3, p4)
if p is None:
continue
# 'probe point' is the most important point which is eventually the camera position (origin).
p = np.array(p)
probe = np.array([p[0], H / 2, p[1]])
# first generate the well-aligned hypothesis.
h = {}
h['roomId'] = room['roomId']
h['type'] = 'wellAligned'
h['probe'] = probe
h['direction'] = -normal3D
h['viewLength'] = np.linalg.norm(middlePoint - p, ord=2)
h['normal'] = normal.copy()
h['wallIndex'] = wallIndex
h['wallJndex'] = wallJndex
h['floorMeta'] = floorMeta
numSeenObjs(room, h, probe, -normal3D, floorMeta, theta)
h['targetWallArea'] = H * np.linalg.norm(
floorMeta[wallIndex][0:2] - floorMeta[wallIndexNext][0:2],
ord=2)
h['targetWallNumWindows'] = targetWallNumWindows
h['targetWallWindoorArea'] = targetWallWindoorArea
# tarWindoorArea2021(h, scene, floorMeta, theta)
h['theta'] = theta
# hypotheses.append(h)
# then we try following the 'Three-Wall' rule. (Left Side)
expandPre, expandNxt = expandWallSeg(wallIndex, floorMeta)
pThW1 = None
pThW2 = None
if expandPre is not None and expandNxt is not None:
pThW1 = twoInfLineIntersection(
expandPre, expandPre + floorMeta[wallIndex][2:4], p3, p4)
pThW2 = twoInfLineIntersection(
expandNxt, expandNxt + floorMeta[wallIndex][2:4], p3, p4)
if pThW1 is not None and pThW2 is not None:
pThW1, pThW2 = np.array(pThW1), np.array(pThW2)
thw = h.copy()
thw['type'] = 'threeWall'
thw['pThW1'] = pThW1
thw['pThW2'] = pThW2
thw['probe'] = pThW1 + (pThW2 - pThW1) / 3
thw['probe'] = np.array(
[thw['probe'][0], H / 2, thw['probe'][1]])
# thw['direction'] = np.array([floorMeta[wallIndexNext][0], H/2, floorMeta[wallIndexNext][1]]) - thw['probe']
# acr = floorMeta[wallIndexNext][0:2] + (floorMeta[wallIndex][0:2] - floorMeta[wallIndexNext][0:2])/3
acr = expandNxt + (expandPre - expandNxt) / 3
thw['direction'] = np.array([acr[0], H / 2, acr[1]
]) - thw['probe']
thw['direction'] /= np.linalg.norm(thw['direction'])
thw['direction'] = groundShifting(thw['probe'], floorMeta,
floorPoly, thw['direction'],
theta, H)
thw['viewLength'] = np.linalg.norm(np.array([
floorMeta[wallIndexNext][0], H / 2,
floorMeta[wallIndexNext][1]
]) - thw['probe'],
ord=2)
hypotheses.append(thw)
# then we try following the 'Three-Wall' rule. (Right Side)
thwR = thw.copy()
thwR['probe'] = pThW2 + (pThW1 - pThW2) / 3
thwR['type'] = 'threeWall_R'
thwR['probe'] = np.array(
[thwR['probe'][0], H / 2, thwR['probe'][1]])
# thwR['direction'] = np.array([floorMeta[wallIndex][0], H/2, floorMeta[wallIndex][1]]) - thwR['probe']
# acr = floorMeta[wallIndex][0:2] + (floorMeta[wallIndexNext][0:2] - floorMeta[wallIndex][0:2])/3
acr = expandPre + (expandNxt - expandPre) / 3
thwR['direction'] = np.array([acr[0], H / 2, acr[1]
]) - thwR['probe']
thwR['direction'] /= np.linalg.norm(thwR['direction'])
thwR['direction'] = groundShifting(thwR['probe'], floorMeta,
floorPoly,
thwR['direction'], theta, H)
thwR['viewLength'] = np.linalg.norm(np.array([
floorMeta[wallIndexNext][0], H / 2,
floorMeta[wallIndexNext][1]
]) - thwR['probe'],
ord=2)
hypotheses.append(thwR)
# then we try following the 'Three-Wall' rule. (Right Side)
mtm = thw.copy()
mtm['probe'] = (pThW1 + pThW2) / 2
mtm['type'] = 'wellAlignedShifted'
mtm['probe'] = np.array(
[mtm['probe'][0], H / 2, mtm['probe'][1]])
# thwR['direction'] = np.array([floorMeta[wallIndex][0], H/2, floorMeta[wallIndex][1]]) - thwR['probe']
# acr = floorMeta[wallIndex][0:2] + (floorMeta[wallIndexNext][0:2] - floorMeta[wallIndex][0:2])/3
acr = (expandNxt + expandPre) / 2
mtm['direction'] = np.array([acr[0], H / 2, acr[1]
]) - mtm['probe']
mtm['direction'] /= np.linalg.norm(mtm['direction'])
mtm['direction'] = groundShifting(mtm['probe'], floorMeta,
floorPoly, mtm['direction'],
theta, H)
mtm['viewLength'] = np.linalg.norm(np.array([
floorMeta[wallIndexNext][0], H / 2,
floorMeta[wallIndexNext][1]
]) - mtm['probe'],
ord=2)
hypotheses.append(mtm)
# the prefix wall and the suffix wall
pThW1 = twoInfLineIntersection(
floorMeta[(wallIndex + floorMeta.shape[0] - 1) %
floorMeta.shape[0]][0:2], floorMeta[wallIndex][0:2],
p3, p4)
pThW2 = twoInfLineIntersection(
floorMeta[wallIndexNext][0:2],
floorMeta[(wallIndexNext + 1) % floorMeta.shape[0]][0:2], p3,
p4)
if pThW1 is not None and pThW2 is not None:
pThW1, pThW2 = np.array(pThW1), np.array(pThW2)
thinL = h.copy()
thinL['type'] = 'threeWall_thin'
thinL['pThW1'] = pThW1
thinL['pThW2'] = pThW2
thinL['probe'] = pThW1 + (pThW2 - pThW1) / 3
thinL['probe'] = np.array(
[thinL['probe'][0], H / 2, thinL['probe'][1]])
acr = floorMeta[wallIndexNext][0:2] + (
floorMeta[wallIndex][0:2] -
floorMeta[wallIndexNext][0:2]) / 3
thinL['direction'] = np.array([acr[0], H / 2, acr[1]
]) - thinL['probe']
thinL['direction'] /= np.linalg.norm(thinL['direction'])
thinL['direction'] = groundShifting(thinL['probe'], floorMeta,
floorPoly,
thinL['direction'], theta,
H)
thinL['viewLength'] = np.linalg.norm(np.array([
floorMeta[wallIndexNext][0], H / 2,
floorMeta[wallIndexNext][1]
]) - thinL['probe'],
ord=2)
hypotheses.append(thinL)
# then we try following the 'Three-Wall' rule. (Right Side)
thinR = thw.copy()
thinR['probe'] = pThW2 + (pThW1 - pThW2) / 3
thinR['type'] = 'threeWall_R_thin'
thinR['probe'] = np.array(
[thinR['probe'][0], H / 2, thinR['probe'][1]])
acr = floorMeta[wallIndex][0:2] + (
floorMeta[wallIndexNext][0:2] -
floorMeta[wallIndex][0:2]) / 3
thinR['direction'] = np.array([acr[0], H / 2, acr[1]
]) - thinR['probe']
thinR['direction'] /= np.linalg.norm(thinR['direction'])
thinR['direction'] = groundShifting(thinR['probe'], floorMeta,
floorPoly,
thinR['direction'], theta,
H)
thinR['viewLength'] = np.linalg.norm(np.array([
floorMeta[wallIndexNext][0], H / 2,
floorMeta[wallIndexNext][1]
]) - thinR['probe'],
ord=2)
hypotheses.append(thinR)
hypotheses = redundancyRemove(hypotheses)
for h in hypotheses:
h['roomTypes'] = room['roomTypes']
h['isObjCovered'] = isObjCovered(h, scene)
theLawOfTheThird(h, room, theta, ASPECT)
numSeenObjs(room, h, h['probe'], h['direction'], floorMeta, theta)
tarWindoorArea2021(h, scene, floorMeta, theta)
layoutConstraint(h, room, theta)
wallNormalOffset(h, floorMeta)
toOriginAndTarget(h)
hypotheses.sort(key=probabilityOPP, reverse=True)
for rank, h in zip(range(0, len(hypotheses)), hypotheses):
h['rank'] = rank
bestViews = {
'wellAlignedShifted': None,
'threeWall_R': None,
'threeWall': None,
'againstMidWall': None,
'twoWallPerspective': None
}
for h in hypotheses:
for viewTps in bestViews:
if viewTps != h['type']:
continue
if bestViews[viewTps] is None:
bestViews[viewTps] = toOriginAndTarget(h)
# bestViews = []
# numOfChosen = min(3, len(hypotheses))
# for index in range(0, numOfChosen):
# h = hypotheses[index]
# bestViews.append(toOriginAndTarget(h))
return hypotheses
def autoViewTwoWallPerspective(room, scene):
fov = scene['PerspectiveCamera']['fov']
# change the fov/2 to Radian.
theta = (np.pi * fov / 180) / 2
focal = 1 / np.tan(theta)
tanPhi = ASPECT / focal
floorMeta = p2d(
'.', '/dataset/room/{}/{}f.obj'.format(room['origin'],
room['modelId']))
floorPoly = Polygon(floorMeta[:, 0:2])
H = sk.getWallHeight(
f"./dataset/room/{room['origin']}/{room['modelId']}w.obj")
pcams = []
for wallDiagIndex in range(floorMeta.shape[0]):
pcam = {}
iPre = (wallDiagIndex + floorMeta.shape[0] - 1) % floorMeta.shape[0]
iNxt = (wallDiagIndex + 1) % floorMeta.shape[0]
iPreP = floorMeta[iPre][0:2]
iNxtP = floorMeta[iNxt][0:2]
# extend two walls as far as possible:
preList = []
nxtList = []
for i in range(floorMeta.shape[0]):
if i == iPre or i == wallDiagIndex:
continue
p3 = floorMeta[i][0:2]
p4 = floorMeta[(i + 1) % floorMeta.shape[0]][0:2]
_p = twoInfLineIntersection(floorMeta[wallDiagIndex][0:2], iPreP,
p3, p4)
if _p is None:
continue
_p = np.array(_p)
if np.dot(_p - floorMeta[wallDiagIndex][0:2],
floorMeta[wallDiagIndex][2:4]) < 0:
continue
preList.append(_p)
for i in range(floorMeta.shape[0]):
if i == iPre or i == wallDiagIndex:
continue
p3 = floorMeta[i][0:2]
p4 = floorMeta[(i + 1) % floorMeta.shape[0]][0:2]
_p = twoInfLineIntersection(floorMeta[wallDiagIndex][0:2], iNxtP,
p3, p4)
if _p is None:
continue
_p = np.array(_p)
if np.dot(_p - floorMeta[wallDiagIndex][0:2],
floorMeta[iPre][2:4]) < 0:
continue
nxtList.append(_p)
MAXdis = -1
for pl in preList:
for nl in nxtList:
if checkPtoN(pl, nl, floorMeta):
dis = np.linalg.norm(pl - nl)
if MAXdis < dis:
MAXdis = dis
iPreP = pl
iNxtP = nl
direction = iNxtP - iPreP
direction = direction[[1, 0]]
direction[1] = -direction[1]
direction /= np.linalg.norm(direction, ord=2)
probe = (iNxtP + iPreP) / 2
if np.dot(direction, floorMeta[wallDiagIndex][0:2] - probe) < 0:
direction = -direction
dis = np.linalg.norm(probe - iNxtP, ord=2) / tanPhi
probe = probe - direction * dis
pcam['viewLength'] = np.linalg.norm(probe -
floorMeta[wallDiagIndex][0:2],
ord=2)
if not floorPoly.contains(Point(probe[0], probe[1])):
p1 = probe
p2 = probe + direction * dis
_plist = []
for i in range(floorMeta.shape[0]):
p3 = floorMeta[i][0:2]
p4 = floorMeta[(i + 1) % floorMeta.shape[0]][0:2]
_p = twoInfLineIntersection(p1, p2, p3, p4)
if _p is None:
continue
if np.dot(direction, np.array(_p) - p2) > 0:
continue
_plist.append(_p)
if len(_plist) > 0:
_i = np.argmin(np.linalg.norm(np.array(_plist), axis=1))
probe = _plist[_i]
else:
probe = probe.tolist()
if not floorPoly.contains(Point(probe[0], probe[1])):
continue
probe.insert(1, H / 2)
probe = np.array(probe)
direction = direction.tolist()
direction.insert(1, 0)
direction = groundShifting(probe, floorMeta, floorPoly,
np.array(direction), theta, H)
pcam['probe'] = probe
pcam['direction'] = direction
pcam['theta'] = theta
pcam['roomId'] = room['roomId']
pcam['wallDiagIndex'] = wallDiagIndex
pcam['type'] = 'twoWallPerspective'
pcam['floorMeta'] = floorMeta
pcams.append(pcam)
return pcams
