socome / BOVW_2019

BOVW(Bag of Visual word)

BOVW(Bag of Visual word) in Colab

Detail

Data Loader 1 (from caltech)

!wget http://www.vision.caltech.edu/Image_Datasets/Caltech101/101_ObjectCategories.tar.gz
!tar -zxvf 101_ObjectCategories.tar.gz

Data Loader 2 (from kaggle)

! kaggle competitions download -c 2019-ml-finalproject
! unzip 2019-ml-finalproject.zip

Codebook(Vocabulary)

(생략)

train_des_arr = np.array(train_des[0])
for remaining in tqdm(train_des[1:]):
  train_des_arr = np.vstack((train_des_arr, remaining))

seeding = kmc2.kmc2(train_des_arr, 200)
Kmeas = MiniBatchKMeans(200, init=seeding).fit(train_des_arr)
scaler = StandardScaler()

Kmeas.cluster_centers_  // Codebook

SPM(Spatial Pyramid Matching)

total_train_feature = list()

for i in tqdm(range(len(train_des))) :

  ## level 0
  level_0 = vq(train_des[i],codebook)
  level_0 = np.histogram(level_0[0],bins=list(range(code_book_size+1)))[0]

  des_map = train_des[i].reshape(des_size,des_size,128)

  ## level 1
  level_1_1 = vq(des_map[:p2,:p2,:].reshape(-1,128),codebook)
  level_1_1_train_feature = np.histogram(level_1_1[0],bins=list(range(code_book_size+1)))
  level_1_1_train_feature = level_1_1_train_feature[0].reshape(code_book_size,1)

  level_1_2 = vq(des_map[p2:,:p2,:].reshape(-1,128),codebook)
  level_1_2_train_feature = np.histogram(level_1_2[0],bins=list(range(code_book_size+1)))
  level_1_2_train_feature = level_1_2_train_feature[0].reshape(code_book_size,1)

  level_1_3 = vq(des_map[:p2,p2:,:].reshape(-1,128),codebook)
  level_1_3_train_feature = np.histogram(level_1_3[0],bins=list(range(code_book_size+1)))
  level_1_3_train_feature = level_1_3_train_feature[0].reshape(code_book_size,1)

  level_1_4 = vq(des_map[p2:,p2:,:].reshape(-1,128),codebook)
  level_1_4_train_feature = np.histogram(level_1_4[0],bins=list(range(code_book_size+1)))
  level_1_4_train_feature = level_1_4_train_feature[0].reshape(code_book_size,1)

  level_1 = np.concatenate((level_1_1_train_feature,level_1_2_train_feature,level_1_3_train_feature,level_1_4_train_feature),axis=1).flatten()

  ## level 2
  level_2_1 = vq(des_map[:p1,:p1,:].reshape(-1,128),codebook)
  level_2_1_train_feature = np.histogram(level_2_1[0],bins=list(range(code_book_size+1)))
  level_2_1_train_feature = level_2_1_train_feature[0].reshape(code_book_size,1)

  level_2_2 = vq(des_map[p1:p2,:p1,:].reshape(-1,128),codebook)
  level_2_2_train_feature = np.histogram(level_2_2[0],bins=list(range(code_book_size+1)))
  level_2_2_train_feature = level_2_2_train_feature[0].reshape(code_book_size,1)

  level_2_3 = vq(des_map[p2:p3,p1:,:].reshape(-1,128),codebook)
  level_2_3_train_feature = np.histogram(level_2_3[0],bins=list(range(code_book_size+1)))
  level_2_3_train_feature = level_2_3_train_feature[0].reshape(code_book_size,1)

  level_2_4 = vq(des_map[p3:,p1:,:].reshape(-1,128),codebook)
  level_2_4_train_feature = np.histogram(level_2_4[0],bins=list(range(code_book_size+1)))
  level_2_4_train_feature = level_2_4_train_feature[0].reshape(code_book_size,1)

  level_2_5 = vq(des_map[:p1,p1:p2,:].reshape(-1,128),codebook)
  level_2_5_train_feature = np.histogram(level_2_5[0],bins=list(range(code_book_size+1)))
  level_2_5_train_feature = level_2_5_train_feature[0].reshape(code_book_size,1)

  level_2_6 = vq(des_map[p1:p2,p1:p2,:].reshape(-1,128),codebook)
  level_2_6_train_feature = np.histogram(level_2_6[0],bins=list(range(code_book_size+1)))
  level_2_6_train_feature = level_2_6_train_feature[0].reshape(code_book_size,1)

  level_2_7 = vq(des_map[p2:p3,p1:p2,:].reshape(-1,128),codebook)
  level_2_7_train_feature = np.histogram(level_2_7[0],bins=list(range(code_book_size+1)))
  level_2_7_train_feature = level_2_7_train_feature[0].reshape(code_book_size,1)

  level_2_8 = vq(des_map[p3:,p1:p2,:].reshape(-1,128),codebook)
  level_2_8_train_feature = np.histogram(level_2_8[0],bins=list(range(code_book_size+1)))
  level_2_8_train_feature = level_2_8_train_feature[0].reshape(code_book_size,1)

  level_2_9 = vq(des_map[:p1,p2:p3,:].reshape(-1,128),codebook)
  level_2_9_train_feature = np.histogram(level_2_9[0],bins=list(range(code_book_size+1)))
  level_2_9_train_feature = level_2_9_train_feature[0].reshape(code_book_size,1)

  level_2_10 = vq(des_map[p1:p2,p2:p3,:].reshape(-1,128),codebook)
  level_2_10_train_feature = np.histogram(level_2_10[0],bins=list(range(code_book_size+1)))
  level_2_10_train_feature = level_2_10_train_feature[0].reshape(code_book_size,1)

  level_2_11 = vq(des_map[p2:p3,p2:p3,:].reshape(-1,128),codebook)
  level_2_11_train_feature = np.histogram(level_2_11[0],bins=list(range(code_book_size+1)))
  level_2_11_train_feature = level_2_11_train_feature[0].reshape(code_book_size,1)

  level_2_12 = vq(des_map[p3:,p2:p3,:].reshape(-1,128),codebook)
  level_2_12_train_feature = np.histogram(level_2_12[0],bins=list(range(code_book_size+1)))
  level_2_12_train_feature = level_2_12_train_feature[0].reshape(code_book_size,1)

  level_2_13 = vq(des_map[:p1,p3:,:].reshape(-1,128),codebook)
  level_2_13_train_feature = np.histogram(level_2_13[0],bins=list(range(code_book_size+1)))
  level_2_13_train_feature = level_2_13_train_feature[0].reshape(code_book_size,1)

  level_2_14 = vq(des_map[p1:p2,p3:,:].reshape(-1,128),codebook)
  level_2_14_train_feature = np.histogram(level_2_14[0],bins=list(range(code_book_size+1)))
  level_2_14_train_feature = level_2_14_train_feature[0].reshape(code_book_size,1)

  level_2_15 = vq(des_map[p2:p3,p3:,:].reshape(-1,128),codebook)
  level_2_15_train_feature = np.histogram(level_2_15[0],bins=list(range(code_book_size+1)))
  level_2_15_train_feature = level_2_15_train_feature[0].reshape(code_book_size,1)

  level_2_16 = vq(des_map[p3:,p3:,:].reshape(-1,128),codebook)
  level_2_16_train_feature = np.histogram(level_2_16[0],bins=list(range(code_book_size+1)))
  level_2_16_train_feature = level_2_16_train_feature[0].reshape(code_book_size,1)

  level_2 = np.concatenate((level_2_1_train_feature,level_2_2_train_feature,level_2_3_train_feature,level_2_4_train_feature,
                            level_2_5_train_feature,level_2_6_train_feature,level_2_7_train_feature,level_2_8_train_feature,
                            level_2_9_train_feature,level_2_10_train_feature,level_2_11_train_feature,level_2_12_train_feature,
                            level_2_13_train_feature,level_2_14_train_feature,level_2_15_train_feature,level_2_16_train_feature),axis=1).flatten()

  total_train_feature.append(np.concatenate((0.25*level_0,0.25*level_1,0.5*level_2)))
  # total_train_feature.append(level_0)

total_train_feature = np.array(total_train_feature)

SPM(Spatial Pyramid Matching) Kernel

def histogramIntersection(M, N):
    m = M.shape[0]
    n = N.shape[0]
    result = np.zeros((m,n))
    for i in range(m):
        for j in range(n):
            temp = np.sum(np.minimum(M[i], N[j]))
            result[i][j] = temp
    return result
    

gramMatrix = histogramIntersection(total_train_feature, total_train_feature)
clf = svm.SVC(kernel='precomputed')
clf.fit(gramMatrix, train_labels)
predictMatrix = histogramIntersection(total_test_feature, total_train_feature)
SVMResults = clf.predict(predictMatrix)

---

Paper

Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories

---

Report

BOVW

Level	codebook_size	step_size	img_size	background	histogram_intersection	scaler	accuracy
0	200	8	256x256	O	-	O	0.41607
0	200	8	256x256	X	-	O	0.43735
2	600	8	256x256	O	X	O	0.50236
2	600	8	256x256	O	O	X	0.57505
2	600	8	256x256	O	O	O	0.58510
2	600	8	256x256	X	O	X	0.60933
2	600	8	256x256	X	O	O	0.62056

VLAD

Level	codebook_size	step_size	img_size	background	accuracy
0	600	8	256x256	X	0.63593

---

Reference

Histogram_intersection

VLAD

---