def saveFacialFeatures(demographicsData):
print("calculating face features")
allData = []
with open(os.path.join(scriptFolder, "US10kData_FULL.p"), "wb") as f:
for i, data in enumerate(demographicsData):
attractiveScore = float(data['Attractive'])
gender = GENDER_DICT[int(data['Gender'])]
if not np.isnan(attractiveScore):
impath = data['Filename']
im = cv2.imread(impath)
try:
landmarks, faceFeatures = getFaceFeatures(im)
mesh, pose, shape_coeffs, blendshape_coeffs = SFM_FACEFITTING.getMeshFromLandmarks(
landmarks, im, num_shape_coefficients_to_fit=10)
facefeaturesCNN = getFaceFeaturesCNN([im], [landmarks])
dataDict = {
"gender": gender,
"attractiveness": attractiveScore,
"landmarks": landmarks,
"facefeatures": faceFeatures,
"facefeatures3D": shape_coeffs,
"pose": pose,
"blendshape_coeffs": blendshape_coeffs,
"impath": impath,
"facefeaturesCNN": facefeaturesCNN,
}
allData.append(dataDict)
pickle.dump(dataDict, f)
print("%i / %i" % (i, len(demographicsData)))
except:
continue
allDataDF = pd.DataFrame(allData)
allDataDF.to_pickle(os.path.join(scriptFolder, "US10kData.p"))
return allDataDF
def getWebResults(im, predictor):
facefeatures_bfm = getFaceFeaturesCNN([im])[0:99]
rating = predictor.predict([facefeatures_bfm])[0]
# new_facefeatures_bfm = facefeatures_bfm.copy()
new_facefeatures_bfm = findBestFeaturesOptimisation(facefeatures_bfm, predictor)
landmarks = getLandmarks(im)
pose_bfm = BFM_FACEFITTING.getPoseFromShapeCeoffs(landmarks, im, facefeatures_bfm)
mesh_bfm = BFM_FACEFITTING.getMeshFromShapeCeoffs(facefeatures_bfm)
modelview, projection, viewport, indexs = decomposePose(mesh_bfm, pose_bfm, im)
results = {}
results["facefeatures"] = facefeatures_bfm.tolist()
results["rating"] = rating
results["new_facefeatures"] = new_facefeatures_bfm.tolist()
results["modelview"] = modelview
results["projection"] = projection
results["viewport"] = viewport
results["indexs"] = indexs
return results
])
max_I = 100
for i, impath in enumerate(glob.glob(srcFolder)):
if i >= max_I:
break
print(impath)
im = cv2.imread(impath)
filename = os.path.basename(impath)
landmarks = getLandmarks(im)
mesh_sfm, pose_sfm, facefeatures_sfm, blendshape_coeffs_sfm = SFM_FACEFITTING.getMeshFromLandmarks(
landmarks, im, num_iterations=300)
# mesh_sfm, pose_sfm, facefeatures_sfm, blendshape_coeffs_sfm = SFM_FACEFITTING.getMeshFromLandmarks(landmarks, im, num_iterations=5000)
facefeatures_bfm = getFaceFeaturesCNN([im])[0:99]
pose_bfm = BFM_FACEFITTING.getPoseFromShapeCeoffs(
landmarks, im, facefeatures_bfm)
mesh_bfm = BFM_FACEFITTING.getMeshFromShapeCeoffs(facefeatures_bfm)
num_coeffs = 12
coeff_deltas = []
for c in range(num_coeffs):
coeffs_up = [0] * num_coeffs
coeffs_dn = [0] * num_coeffs
coeffs_up[c] = 1
coeffs_dn[c] = -1
coeff_deltas.append(coeffs_up)
coeff_deltas.append(coeffs_dn)
# for coeff_delta_vec in coeff_deltas:
#
# gradientX, gradientY = np.gradient(y_preds)
# plt.figure()
# plt.title('svms gradient')
# Q = plt.quiver(X, Y, gradientX, gradientY, units='width')#
# plt.show()
# impath = r"E:\Facedata\RateMe\21_F_423ekl\HeC768w.jpg"
# impath = r"E:\Facedata\RateMe\21_F_42mxvl\fWIPj9e.jpg"
impath = r"E:\Facedata\RateMe\21_M_5mx3xt\a2Odd4w.jpg"
im = cv2.imread(impath)
landmarks = getLandmarks(im)
# mesh_sfm, pose_sfm, shape_coeffs_sfm, blendshape_coeffs_sfm = SFM_FACEFITTING.getMeshFromLandmarks(landmarks, im, num_iterations=300)
shape_coeffs_bfm = getFaceFeaturesCNN([im])[0:99]
pose_bfm = BFM_FACEFITTING.getPoseFromShapeCeoffs(landmarks, im,
shape_coeffs_bfm)
mesh_bfm = BFM_FACEFITTING.getMeshFromShapeCeoffs(shape_coeffs_bfm)
# new_shape_coeffs = np.array(shape_coeffs_sfm).copy()
new_shape_coeffs = np.array(shape_coeffs_bfm).copy()
distance_moved = 0
target_distance = 2
poss = []
while distance_moved < target_distance:
pos = new_shape_coeffs[:2]
v = vectorAtPos(vectors, pos)
v *= 0.1
dstFolder = "./results/"
impath = r"E:\Facedata\RateMe\21_F_423ekl\HeC768w.jpg"
impath = r"C:\Users\ellio\PycharmProjects\circlelines\Beautifier\face3D\me\0132.jpg"
# impath = r"E:\Facedata\RateMe\18_M_2wbhfw\n5DesJ4.jpg"
im = cv2.imread(impath)
filename = os.path.basename(impath)
outfile_bfm = os.path.join(dstFolder, "%s_bfm.ply" % filename)
outfile_sfm_orig = os.path.join(dstFolder, "%s_sfm_orig.ply" % filename)
outfile_com_orig = os.path.join(dstFolder, "%s_com_orig.ply" % filename)
outfile_sfm_opt = os.path.join(dstFolder, "%s_sfm_opt.ply" % filename)
outfile_com_opt = os.path.join(dstFolder, "%s_com_opt.ply" % filename)
#bfm model
bfm_features = getFaceFeaturesCNN([im])
S, T = utils.projectBackBFM(BFM_MODEL, bfm_features)
utils.write_ply(outfile_bfm, S, T, BFM_FACES)
print(bfm_features)
# sfm model
landmarks = getLandmarks(im)
sfm_mesh, pose, sfm_features, blendshape_coeffs = SFM_FACEFITTING.getMeshFromLandmarks(landmarks, im)
sfm_features = np.array(sfm_features)
print(sfm_features)
verts_sfm = np.array(sfm_mesh.vertices)[:, :3]
verts_sfm = verts_sfm + np.array([0, 0, np.max(S[:, 2]) - np.max(verts_sfm[:, 2])])
text_sfm = np.ones_like(verts_sfm) * 128
faces_sfm = np.array(sfm_mesh.tvi)
utils.write_ply(outfile_sfm_orig, verts_sfm, text_sfm, faces_sfm)
def saveFacialFeatures(combinedcsvpath):
df = pd.read_csv(combinedcsvpath)
df.drop('Rating Text', 1)
df.drop('Submission Title', 1)
#filter out the weird ages
filtered = df.loc[df['Submission Age'] >= 18]
filtered = filtered.loc[filtered['Submission Age'] < 50]
grouped = filtered.groupby([
'Submission Id', "Submission Created UTC", "Folder", "Submission Age",
"Submission Gender"
])
sfile = open(os.path.join(scriptFolder, "RateMeData_FULL.p"), "wb")
pickler = pickle.Pickler(sfile)
allData = []
for i, (uniques, group) in enumerate(grouped):
submissionId = uniques[0]
submissionCreatedUTC = uniques[1]
folder = uniques[2]
submissionAge = uniques[3]
submissionGender = uniques[4]
ratings = np.array(group["Rating"].as_matrix().tolist())
ratingDecimal = np.array(group["Decimal"].as_matrix().tolist())
ratingAuthor = np.array(group["Rating Author"].as_matrix().tolist())
ratingPostedUTC = np.array(
group["Rating Posted UTC"].as_matrix().tolist())
ratingText = np.array(group["Rating Text"].as_matrix().tolist())
comments = [(ratingAuthor[k], ratingPostedUTC[k], ratings[k],
ratingDecimal[k], ratingText[k])
for k in range(len(ratings))]
nonEmptyRatings = [
ratings[k] for k in range(len(ratings))
if not math.isnan(ratings[k])
]
_, attractiveness_mean, attractiveness_std = scipy.stats.t.fit(
nonEmptyRatings)
# cleanedRatings = reject_outliers(np.array(nonEmptyRatings))
# attractiveness_mean = np.mean(cleanedRatings)
# attractiveness_std = np.std(cleanedRatings)
#get the image files:
types = ('*.jpg', '*.png', '*.bmp')
impaths = []
for type in types:
impaths.extend(glob.glob(os.path.join(folder, type)))
usedImPaths = []
ims = []
landmarkss = []
faceFeaturess = []
numBodyShots = 0
numBuddyShots = 0
for impath in impaths:
im = cv2.imread(impath)
# im = ensureImageLessThanMax(im)
try:
landmarks, faceFeatures = getFaceFeatures(im)
usedImPaths.append(impath)
ims.append(im)
landmarkss.append(landmarks)
faceFeaturess.append(faceFeatures)
except Exception as e:
if "No face" in str(e):
numBodyShots += 1
elif "Multiple faces" in str(e):
numBuddyShots += 1
continue
if len(ims) > 0:
print("(%i/%i) #i:%i" % (i, len(grouped), len(ims)))
print(usedImPaths)
meshs, poses, shape_coeffs, blendshape_coeffss = SFM_FACEFITTING.getMeshFromMultiLandmarks(
landmarkss,
ims,
num_shape_coefficients_to_fit=-1,
num_iterations=300)
facefeaturesCNN = getFaceFeaturesCNN(ims, landmarkss)
dataDict = {
"submissionId": submissionId,
"submissionCreatedUTC": submissionCreatedUTC,
"gender": submissionGender,
"age": submissionAge,
"comments": comments,
"ratings": nonEmptyRatings,
"attractiveness": attractiveness_mean,
"attractiveness_std": attractiveness_std,
"numUsableImages": len(ims),
"numSubmittedImages": len(impaths),
"numBodyShots": numBodyShots,
"numBuddyShots": numBuddyShots,
"impaths": usedImPaths,
"landmarkss": landmarkss,
"facefeaturess": faceFeaturess,
# "meshs": meshs, can be regenerated from the coeffs below
"poses": poses,
"facefeatures3D": shape_coeffs,
"blendshape_coeffss": blendshape_coeffss,
"facefeaturesCNN": facefeaturesCNN,
}
allData.append(dataDict)
pickler.dump(dataDict)
sfile.close()
allDataDF = pd.DataFrame(allData)
allDataDF = allDataDF.sample(frac=1).reset_index(drop=True)
allDataDF.to_pickle(os.path.join(scriptFolder, "RateMeData.p"))
return allDataDF