train_des = []
for img in tqdm(train_imgs):
kp, des = dense_sift_each(py0)
train_des.extend(des)
print(type(des))
num_centers = 1024
dataset = torch.from_numpy(des).to(device_gpu)
print('Starting clustering')
centers, codes = cluster(dataset, num_centers)
print(codes)
codebook = centers.cpu().detach().numpy()
his = build(train_imgs, codebook)
x_data = []
for data in his:
py = pyramid(data, 2)
x_data.append(py)
his_test = build(test_imgs, codebook)
x_test = []
for data in his_test:
py = pyramid(data, 2)
x_test.append(py)
x_train = np.asarray(x_data, dtype=np.float32)
x_test = np.asarray(x_test, dtype=np.float32)
x_train_t, x_test_t, y_train_t, y_test_t = train_test_split(x_train, y_train)
result = HI_SVM(x_train_t, x_test_t, y_train_t)
#result = LSVM(x_train_t, x_test_t, y_train_t)
print(classification_report(y_test_t, result))
DataSet : Caltech 101 Scaler : Srandard 적용 각 train 30장을 75:25 비율로 split해서 테스트를 진행
| Level | CodeBook_Size | Kernel | Accuracy |
|---|---|---|---|
| 0 | 256 | Linear | 0.31 |
| 0 | 256 | HI | 0.41 |
| 1 | 256 | Linear | 0.44 |
| 1 | 256 | HI | 0.49 |
| 2 | 256 | Linear | 0.43 |
| 2 | 256 | HI | 0.56 |
| 0 | 512 | Linear | 0.39 |
| 0 | 512 | HI | 0.46 |
| 1 | 512 | Linear | 0.50 |
| 1 | 512 | HI | 0.53 |
| 2 | 512 | Linear | 0.56 |
| 2 | 512 | HI | 0.60 |
| 0 | 2048 | Linear | 0.43 |
| 0 | 2048 | HI | 0.49 |
| 1 | 2048 | Linear | 0.52 |
| 1 | 2048 | HI | 0.59 |
| 2 | 2048 | Linear | 0.58 |
| 2 | 2048 | HI | 0.62 |
코드북 사이즈 : 2048 아래 그림의 정보를 토대로 2048 * 21 = 43,008을 선정해서 튜닝함
VQ = Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories 출처
def HistogramIntersection(X, Y):
x = X.shape[0]
y = Y.shape[0]
result = np.zeros((x,y))
for i in range(x):
for j in range(y):
temp = np.sum(np.minimum(X[i], Y[j]))
result[i][j] = temp
return result
def HI_SVM(x_train, x_test, y_train):
gramMatrix = HistogramIntersection(x_train, x_train)
clf = SVC(kernel='precomputed')
clf.fit(gramMatrix, y_train)
predictMatrix = HistogramIntersection(x_test, x_train)
SVMResults = clf.predict(predictMatrix)
return SVMResults
기술자를 ((이미지 높이 h) / Step size(dense sift), (이미지 높이 h) / Step size(dense sift), 128)로 resize resize된 기술자를 16등분 16등분된 Histogram을 Pyramid 레벨에 맞게 재조합
def build(imgs, codebook):
data = []
for img in tqdm(imgs):
_, des = dense_sift_each(img)
des = np.asarray(des)
des = np.resize(des, (32, 32, 128))
des = cutted(des, 2)
des = np.asarray(des)
l, _, _, _ = des.shape
hist = []
for i in range(0,l):
tmp = np.resize(des[i], (64 ,128))
his = histogram(tmp, codebook)
hist.append(his)
data.append(hist)
return data
16등분을 위한 함수 (이전 kaggle Discussion에서 공개한 피라미드)
def cutted(src, level=2):
h_end, w_end, _ = src.shape
cutted_img = []
w_start = 0
h_start = 0
w = w_end // (2**level)
h = h_end // (2**level)
if level != 0:
for j in range(2 ** level):
for i in range(2 ** level):
img = src[h_start:h_start + h, w_start:w_start + w]
cutted_img.append(img)
w_start += h
if (w_start+h == w_end+w):
w_start = 0
h_start += h
else: cutted_img = src
return cutted_img
def pyramid(his, level):
his1 = np.array([[his[0] + his[1] + his[4] + his[5]], [his[2] + his[3] + his[6] + his[7]], [his[8] + his[9] + his[12] + his[13]], [his[10] + his[11] + his[14] + his[15]]])
his0 = his1[0] + his1[1] + his1[2] + his1[3]
his0 = np.ravel(his0, order='C')
his1 = np.ravel(his1, order='C')
his2 = his
his2 = np.ravel(his2, order='C')
if level == 0:
return np.asarray(his0)
py = np.hstack((his0 * 0.25, his1 * 0.25))
if level == 1:
return py
if level == 2:
return np.hstack((py, his2 * 0.5))