def filter_reconstruct(src, thd=0.3):
src = os.path.abspath(src)
res = {}
if os.path.isdir(src + '/opensfm'): # for odm
cam = src + '/cameras.json'
src += '/opensfm'
rec = src + '/reconstruction.topocentric.json'
elif os.path.isfile(src + '/../cameras.json'):
cam = src + '/../cameras.json' # for odm
rec = src + '/reconstruction.topocentric.json'
elif os.path.isfile(src + '/camera_models.json'):
cam = src + '/camera_models.json' # for sfm
rec = src + '/reconstruction.json'
cam = Camera(cam)
bak = rec[:-4] + 'bak'
if os.path.isfile(bak):
if os.path.isfile(rec): os.remove(rec)
os.rename(bak, rec) # for win
with open(rec) as f:
data = json.load(f)[0]
os.rename(rec, bak)
INFO(f'Filter: {rec}')
from opensfm.dataset import DataSet
T = DataSet(src).load_tracks_manager()
for im in T.get_shot_ids():
v = data['shots'][im]
rotation = np.array(v['rotation'])
translation = np.array(v['translation'])
O, X, Y, Z = calc_axis_cam(translation, rotation)
feat = load_feature(f'{src}/features/{im}', cam, 1)
for tid, x in T.get_shot_observations(im).items():
if not tid in data['points']: continue
dp = data['points'][tid]['coordinates'] - O
ddp = np.linalg.norm(dp)
u, v = feat[x.id][:2] # fid
qt = u * X + v * Y + Z
qt /= np.linalg.norm(qt)
delta = np.dot(dp, qt)
dis = np.sqrt(ddp**2 - delta**2)
if tid not in res: res[tid] = dis
elif dis > res[tid]: res[tid] = dis # meters
#print(f'{im} %6s %6s %.3f'%(tid,x.id,dis))
dis = [*res.values()]
md = np.mean(dis)
thd = min(thd, md)
out = {k: v
for k, v in res.items() if v > thd}
#print(out)
#plt.hist(dis, [0.01,0.05,0.1,0.5,1,2]); plt.show()
for tid in out:
data['points'].pop(tid)
with open(rec, 'w') as f:
json.dump([data], f, indent=4)
INFO('Out=%d/%d, Thd=%.3f, Max=%.3f' % (len(out), len(res), thd, max(dis)))
def SfM_parse_track(src):
from opensfm.dataset import DataSet
#from opensfm.pymap import Observation
TM = DataSet(src).load_tracks_manager()
T = {}
for im in TM.get_shot_ids():
T.setdefault(im, [[], [], [], []]) # tid=str
for tid, v in TM.get_shot_observations(im).items():
T[im][0] += [tid]
T[im][1] += [v.id] # xys
T[im][2] += [np.hstack([v.point, v.scale])]
T[im][3] += [v.color] # RGB
return T # {im: [tid,fid,xys,RGB]}
