基于Pytorch的從零開始的目標(biāo)檢測 | 附源碼
引言
目標(biāo)檢測是計算機(jī)視覺中一個非常流行的任務(wù),在這個任務(wù)中,給定一個圖像,你預(yù)測圖像中物體的包圍盒(通常是矩形的) ,并且識別物體的類型。在這個圖像中可能有多個對象,而且現(xiàn)在有各種先進(jìn)的技術(shù)和框架來解決這個問題,例如 Faster-RCNN 和 YOLOv3。
本文將討論圖像中只有一個感興趣的對象的情況。這里的重點更多是關(guān)于如何讀取圖像及其邊界框、調(diào)整大小和正確執(zhí)行增強(qiáng),而不是模型本身。目標(biāo)是很好地掌握對象檢測背后的基本思想,你可以對其進(jìn)行擴(kuò)展以更好地理解更復(fù)雜的技術(shù)。
本文中的所有代碼都在下面的鏈接中:https://jovian.ai/aakanksha-ns/road-signs-bounding-box-prediction。
問題陳述
給定一個由路標(biāo)組成的圖像,預(yù)測路標(biāo)周圍的包圍盒,并識別路標(biāo)的類型。這些路標(biāo)包括以下四種:
- 紅綠燈
- 停止
- 車速限制
- 人行橫道
這就是所謂的多任務(wù)學(xué)習(xí)問題,因為它涉及執(zhí)行兩個任務(wù): 1)回歸找到包圍盒坐標(biāo),2)分類識別道路標(biāo)志的類型。
1.數(shù)據(jù)集
我使用了來自 Kaggle 的道路標(biāo)志檢測數(shù)據(jù)集,鏈接如下:https://www.kaggle.com/andrewmvd/road-sign-detection
它由877張圖像組成。這是一個相當(dāng)不平衡的數(shù)據(jù)集,大多數(shù)圖像屬于限速類,但由于我們更關(guān)注邊界框預(yù)測,因此可以忽略不平衡。
2.加載數(shù)據(jù)
每個圖像的注釋都存儲在單獨的 XML 文件中。我按照以下步驟創(chuàng)建了訓(xùn)練數(shù)據(jù)集:
- 遍歷訓(xùn)練目錄以獲得所有.xml 文件的列表。
- 使用xml.etree.ElementTree解析.xml文件。
- 創(chuàng)建一個由文件路徑、寬度、高度、邊界框坐標(biāo)( xmin 、 xmax 、 ymin 、 ymax )和每個圖像的類組成的字典,并將字典附加到列表中。
- 使用圖像統(tǒng)計數(shù)據(jù)字典列表創(chuàng)建一個 Pandas 數(shù)據(jù)庫。
def filelist(root, file_type):
"""Returns a fully-qualified list of filenames under root directory"""
return [os.path.join(directory_path, f) for directory_path, directory_name,
files in os.walk(root) for f in files if f.endswith(file_type)]
def generate_train_df (anno_path):
annotations = filelist(anno_path, '.xml')
anno_list = []
for anno_path in annotations:
root = ET.parse(anno_path).getroot()
anno = {}
anno['filename'] = Path(str(images_path) + '/'+ root.find("./filename").text)
anno['width'] = root.find("./size/width").text
anno['height'] = root.find("./size/height").text
anno['class'] = root.find("./object/name").text
anno['xmin'] = int(root.find("./object/bndbox/xmin").text)
anno['ymin'] = int(root.find("./object/bndbox/ymin").text)
anno['xmax'] = int(root.find("./object/bndbox/xmax").text)
anno['ymax'] = int(root.find("./object/bndbox/ymax").text)
anno_list.append(anno)
return pd.DataFrame(anno_list)
- 標(biāo)簽編碼類列
#label encode target
class_dict = {'speedlimit': 0, 'stop': 1, 'crosswalk': 2, 'trafficlight': 3}
df_train['class'] = df_train['class'].apply(lambda x: class_dict[x])
3.調(diào)整圖像和邊界框的大小
由于訓(xùn)練一個計算機(jī)視覺模型需要的圖像是相同的大小,我們需要調(diào)整我們的圖像和他們相應(yīng)的包圍盒。調(diào)整圖像的大小很簡單,但是調(diào)整包圍盒的大小有點棘手,因為每個包圍盒都與圖像及其尺寸相關(guān)。
下面是調(diào)整包圍盒大小的工作原理:
- 將邊界框轉(zhuǎn)換為與其對應(yīng)的圖像大小相同的圖像(稱為掩碼)。這個掩碼只有 0 表示背景,1 表示邊界框覆蓋的區(qū)域。
- 將掩碼調(diào)整到所需的尺寸。
從調(diào)整完大小的掩碼中提取邊界框坐標(biāo)。
def create_mask(bb, x):
"""Creates a mask for the bounding box of same shape as image"""
rows,cols,*_ = x.shape
Y = np.zeros((rows, cols))
bb = bb.astype(np.int)
Y[bb[0]:bb[2], bb[1]:bb[3]] = 1.
return Y
def mask_to_bb(Y):
"""Convert mask Y to a bounding box, assumes 0 as background nonzero object"""
cols, rows = np.nonzero(Y)
if len(cols)==0:
return np.zeros(4, dtype=np.float32)
top_row = np.min(rows)
left_col = np.min(cols)
bottom_row = np.max(rows)
right_col = np.max(cols)
return np.array([left_col, top_row, right_col, bottom_row], dtype=np.float32)
def create_bb_array(x):
"""Generates bounding box array from a train_df row"""
return np.array([x[5],x[4],x[7],x[6]])
def resize_image_bb(read_path,write_path,bb,sz):
"""Resize an image and its bounding box and write image to new path"""
im = read_image(read_path)
im_resized = cv2.resize(im, (int(1.49*sz), sz))
Y_resized = cv2.resize(create_mask(bb, im), (int(1.49*sz), sz))
new_path = str(write_path/read_path.parts[-1])
cv2.imwrite(new_path, cv2.cvtColor(im_resized, cv2.COLOR_RGB2BGR))
return new_path, mask_to_bb(Y_resized)
#Populating Training DF with new paths and bounding boxes
new_paths = []
new_bbs = []
train_path_resized = Path('./road_signs/images_resized')
for index, row in df_train.iterrows():
new_path,new_bb = resize_image_bb(row['filename'], train_path_resized, create_bb_array(row.values),300)
new_paths.append(new_path)
new_bbs.append(new_bb)
df_train['new_path'] = new_paths
df_train['new_bb'] = new_bbs
4.數(shù)據(jù)增強(qiáng)
數(shù)據(jù)增強(qiáng)是一種通過使用現(xiàn)有圖像的不同變體創(chuàng)建新的訓(xùn)練圖像來更好地概括我們的模型的技術(shù)。我們當(dāng)前的訓(xùn)練集中只有 800 張圖像,因此數(shù)據(jù)增強(qiáng)對于確保我們的模型不會過擬合非常重要。
對于這個問題,我使用了翻轉(zhuǎn)、旋轉(zhuǎn)、中心裁剪和隨機(jī)裁剪。
這里唯一需要記住的是確保包圍盒也以與圖像相同的方式進(jìn)行轉(zhuǎn)換。
# modified from fast.ai
def crop(im, r, c, target_r, target_c):
return im[r:r+target_r, c:c+target_c]
# random crop to the original size
def random_crop(x, r_pix=8):
""" Returns a random crop"""
r, c,*_ = x.shape
c_pix = round(r_pix*c/r)
rand_r = random.uniform(0, 1)
rand_c = random.uniform(0, 1)
start_r = np.floor(2*rand_r*r_pix).astype(int)
start_c = np.floor(2*rand_c*c_pix).astype(int)
return crop(x, start_r, start_c, r-2*r_pix, c-2*c_pix)
def center_crop(x, r_pix=8):
r, c,*_ = x.shape
c_pix = round(r_pix*c/r)
return crop(x, r_pix, c_pix, r-2*r_pix, c-2*c_pix)
def rotate_cv(im, deg, y=False, mode=cv2.BORDER_REFLECT, interpolation=cv2.INTER_AREA):
""" Rotates an image by deg degrees"""
r,c,*_ = im.shape
M = cv2.getRotationMatrix2D((c/2,r/2),deg,1)
if y:
return cv2.warpAffine(im, M,(c,r), borderMode=cv2.BORDER_CONSTANT)
return cv2.warpAffine(im,M,(c,r), borderMode=mode, flags=cv2.WARP_FILL_OUTLIERS+interpolation)
def random_cropXY(x, Y, r_pix=8):
""" Returns a random crop"""
r, c,*_ = x.shape
c_pix = round(r_pix*c/r)
rand_r = random.uniform(0, 1)
rand_c = random.uniform(0, 1)
start_r = np.floor(2*rand_r*r_pix).astype(int)
start_c = np.floor(2*rand_c*c_pix).astype(int)
xx = crop(x, start_r, start_c, r-2*r_pix, c-2*c_pix)
YY = crop(Y, start_r, start_c, r-2*r_pix, c-2*c_pix)
return xx, YY
def transformsXY(path, bb, transforms):
x = cv2.imread(str(path)).astype(np.float32)
x = cv2.cvtColor(x, cv2.COLOR_BGR2RGB)/255
Y = create_mask(bb, x)
if transforms:
rdeg = (np.random.random()-.50)*20
x = rotate_cv(x, rdeg)
Y = rotate_cv(Y, rdeg, y=True)
if np.random.random() > 0.5:
x = np.fliplr(x).copy()
Y = np.fliplr(Y).copy()
x, Y = random_cropXY(x, Y)
else:
x, Y = center_crop(x), center_crop(Y)
return x, mask_to_bb(Y)
def create_corner_rect(bb, color='red'):
bb = np.array(bb, dtype=np.float32)
return plt.Rectangle((bb[1], bb[0]), bb[3]-bb[1], bb[2]-bb[0], color=color,
fill=False, lw=3)
def show_corner_bb(im, bb):
plt.imshow(im)
plt.gca().add_patch(create_corner_rect(bb))
圖片
5.PyTorch 數(shù)據(jù)集
現(xiàn)在我們已經(jīng)有了數(shù)據(jù)增強(qiáng),我們可以進(jìn)行訓(xùn)練驗證拆分并創(chuàng)建我們的 PyTorch 數(shù)據(jù)集。我們使用 ImageNet 統(tǒng)計數(shù)據(jù)對圖像進(jìn)行標(biāo)準(zhǔn)化,因為我們使用的是預(yù)訓(xùn)練的 ResNet 模型并在訓(xùn)練時在我們的數(shù)據(jù)集中應(yīng)用數(shù)據(jù)增強(qiáng)。
X_train, X_val, y_train, y_val = train_test_split(X, Y, test_size=0.2, random_state=42)
def normalize(im):
"""Normalizes images with Imagenet stats."""
imagenet_stats = np.array([[0.485, 0.456, 0.406], [0.229, 0.224, 0.225]])
return (im - imagenet_stats[0])/imagenet_stats[1]
class RoadDataset(Dataset):
def __init__(self, paths, bb, y, transforms=False):
self.transforms = transforms
self.paths = paths.values
self.bb = bb.values
self.y = y.values
def __len__(self):
return len(self.paths)
def __getitem__(self, idx):
path = self.paths[idx]
y_class = self.y[idx]
x, y_bb = transformsXY(path, self.bb[idx], self.transforms)
x = normalize(x)
x = np.rollaxis(x, 2)
return x, y_class, y_bb
train_ds = RoadDataset(X_train['new_path'],X_train['new_bb'] ,y_train, transforms=True)
valid_ds = RoadDataset(X_val['new_path'],X_val['new_bb'],y_val)
batch_size = 64
train_dl = DataLoader(train_ds, batch_size=batch_size, shuffle=True)
valid_dl = DataLoader(valid_ds, batch_size=batch_size)
6.PyTorch 模型
對于這個模型,我使用了一個非常簡單的預(yù)先訓(xùn)練的 resNet-34模型。由于我們有兩個任務(wù)要完成,這里有兩個最后的層: 包圍盒回歸器和圖像分類器。
class BB_model(nn.Module):
def __init__(self):
super(BB_model, self).__init__()
resnet = models.resnet34(pretrained=True)
layers = list(resnet.children())[:8]
self.features1 = nn.Sequential(*layers[:6])
self.features2 = nn.Sequential(*layers[6:])
self.classifier = nn.Sequential(nn.BatchNorm1d(512), nn.Linear(512, 4))
self.bb = nn.Sequential(nn.BatchNorm1d(512), nn.Linear(512, 4))
def forward(self, x):
x = self.features1(x)
x = self.features2(x)
x = F.relu(x)
x = nn.AdaptiveAvgPool2d((1,1))(x)
x = x.view(x.shape[0], -1)
return self.classifier(x), self.bb(x)
7.訓(xùn)練
對于損失,我們需要同時考慮分類損失和邊界框回歸損失,因此我們使用交叉熵和 L1 損失(真實值和預(yù)測坐標(biāo)之間的所有絕對差之和)的組合。我已經(jīng)將 L1 損失縮放了 1000 倍,因為分類和回歸損失都在相似的范圍內(nèi)。除此之外,它是一個標(biāo)準(zhǔn)的 PyTorch 訓(xùn)練循環(huán)(使用 GPU):
def update_optimizer(optimizer, lr):
for i, param_group in enumerate(optimizer.param_groups):
param_group["lr"] = lr
def train_epocs(model, optimizer, train_dl, val_dl, epochs=10,C=1000):
idx = 0
for i in range(epochs):
model.train()
total = 0
sum_loss = 0
for x, y_class, y_bb in train_dl:
batch = y_class.shape[0]
x = x.cuda().float()
y_class = y_class.cuda()
y_bb = y_bb.cuda().float()
out_class, out_bb = model(x)
loss_class = F.cross_entropy(out_class, y_class, reduction="sum")
loss_bb = F.l1_loss(out_bb, y_bb, reduction="none").sum(1)
loss_bb = loss_bb.sum()
loss = loss_class + loss_bb/C
optimizer.zero_grad()
loss.backward()
optimizer.step()
idx += 1
total += batch
sum_loss += loss.item()
train_loss = sum_loss/total
val_loss, val_acc = val_metrics(model, valid_dl, C)
print("train_loss %.3f val_loss %.3f val_acc %.3f" % (train_loss, val_loss, val_acc))
return sum_loss/total
def val_metrics(model, valid_dl, C=1000):
model.eval()
total = 0
sum_loss = 0
correct = 0
for x, y_class, y_bb in valid_dl:
batch = y_class.shape[0]
x = x.cuda().float()
y_class = y_class.cuda()
y_bb = y_bb.cuda().float()
out_class, out_bb = model(x)
loss_class = F.cross_entropy(out_class, y_class, reduction="sum")
loss_bb = F.l1_loss(out_bb, y_bb, reduction="none").sum(1)
loss_bb = loss_bb.sum()
loss = loss_class + loss_bb/C
_, pred = torch.max(out_class, 1)
correct += pred.eq(y_class).sum().item()
sum_loss += loss.item()
total += batch
return sum_loss/total, correct/total
model = BB_model().cuda()
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adam(parameters, lr=0.006)
train_epocs(model, optimizer, train_dl, valid_dl, epochs=15)
8.測試
現(xiàn)在我們已經(jīng)完成了訓(xùn)練,我們可以選擇一個隨機(jī)圖像并在上面測試我們的模型。盡管我們只有相當(dāng)少量的訓(xùn)練圖像,但是我們最終在測試圖像上得到了一個相當(dāng)不錯的預(yù)測。
使用手機(jī)拍攝真實照片并測試模型將是一項有趣的練習(xí)。另一個有趣的實驗是不執(zhí)行任何數(shù)據(jù)增強(qiáng)并訓(xùn)練模型并比較兩個模型。
# resizing test image
im = read_image('./road_signs/images_resized/road789.png')
im = cv2.resize(im, (int(1.49*300), 300))
cv2.imwrite('./road_signs/road_signs_test/road789.jpg', cv2.cvtColor(im, cv2.COLOR_RGB2BGR))
# test Dataset
test_ds = RoadDataset(pd.DataFrame([{'path':'./road_signs/road_signs_test/road789.jpg'}])['path'],pd.DataFrame([{'bb':np.array([0,0,0,0])}])['bb'],pd.DataFrame([{'y':[0]}])['y'])
x, y_class, y_bb = test_ds[0]
xx = torch.FloatTensor(x[None,])
xx.shape
# prediction
out_class, out_bb = model(xx.cuda())
out_class, out_bb
總結(jié)
現(xiàn)在我們已經(jīng)介紹了目標(biāo)檢測的基本原理,并從頭開始實現(xiàn)它,您可以將這些想法擴(kuò)展到多對象情況,并嘗試更復(fù)雜的模型,如 RCNN 和 YOLO!