.

Author: cs-giung

demo_detect.py

@@ -40,21 +40,21 @@ def plot_figures(figures, nrows=1, ncols=1):
 
 # S3FD returns bboxes.
 t = time.time()
-bboxes = DET3.detect_faces(img, conf_th=0.8, scales=[0.5])
+bboxes = DET3.detect_faces(img, conf_th=0.9, scales=[0.5])
 print('S3FD : %d faces in %.4f seconds.' % (len(bboxes), time.time() - t))
 img3 = draw_bboxes(img, bboxes)
 
 # DSFD returns bboxes.
 t = time.time()
-bboxes = DET4.detect_faces(img, conf_th=0.8, scales=[0.5])
+bboxes = DET4.detect_faces(img, conf_th=0.9, scales=[0.5])
 print('DSFD : %d faces in %.4f seconds.' % (len(bboxes), time.time() - t))
 img4 = draw_bboxes(img, bboxes)
 
 # plot results.
 results = {
-    'MTCNN (conf_th=0.9, scales=[1])': img1,
+    'MTCNN (conf_th=0.9, scales=[0.125])': img1,
     'Tiny Face (conf_th=0.9, scales=[1])': img2,
-    'S3FD (conf_th=0.9, scales=[1])': img3,
-    'DSFD (conf_th=0.9, scales=[1])': img4
+    'S3FD (conf_th=0.9, scales=[0.5])': img3,
+    'DSFD (conf_th=0.9, scales=[0.5])': img4
 }
 plot_figures(results, 2, 2)

detectors/__init__.py

@@ -1,4 +1,4 @@
 from .mtcnn import MTCNN
 from .tinyface import TinyFace
 from .s3fd import S3FD
-from .dsfd import DSFD, DSFD_
+from .dsfd import DSFD

detectors/dsfd/__init__.py

@@ -4,11 +4,9 @@
 import torch
 from torchvision import transforms
 from .nets import DSFDNet
-from .nets2 import DSFDNet2
 from .box_utils import nms_
 
 PATH_WEIGHT = './detectors/dsfd/weights/dsfd_vgg_0.880.pth'
-PATH_WEIGHT2 = './detectors/dsfd/weights/WIDERFace_DSFD_RES152.pth'
 img_mean = np.array([104., 117., 123.])[:, np.newaxis, np.newaxis].astype('float32')
 
 
@@ -61,51 +59,3 @@ def detect_faces(self, image, conf_th=0.8, scales=[1]):
             bboxes = bboxes[keep]
 
         return bboxes
-
-
-class DSFD_():
-
-    def __init__(self, device='cuda'):
-
-        tstamp = time.time()
-        self.device = device
-
-        print('[DSFD] loading with', self.device)
-        self.net = DSFDNet2(device=self.device).to(self.device)
-        state_dict = torch.load(PATH_WEIGHT2, map_location=self.device)
-        self.net.load_state_dict(state_dict)
-        self.net.eval()
-        print('[DSFD] finished loading (%.4f sec)' % (time.time() - tstamp))
-    
-    def detect_faces(self, image, conf_th=0.8, scales=[1]):
-
-        w, h = image.shape[1], image.shape[0]
-
-        bboxes = np.empty(shape=(0, 5))
-
-        with torch.no_grad():
-            for s in scales:
-                scaled_img = cv2.resize(image, dsize=(0, 0), fx=s, fy=s, interpolation=cv2.INTER_LINEAR)
-
-                scaled_img = np.swapaxes(scaled_img, 1, 2)
-                scaled_img = np.swapaxes(scaled_img, 1, 0)
-                scaled_img = scaled_img[[2, 1, 0], :, :]
-                scaled_img = scaled_img.astype('float32')
-                scaled_img -= img_mean
-                scaled_img = scaled_img[[2, 1, 0], :, :]
-                x = torch.from_numpy(scaled_img).unsqueeze(0).to(self.device)
-                y = self.net(x)
-
-                detections = y.data
-                scale = torch.Tensor([w, h, w, h])
-
-                for i in range(detections.size(1)):
-                    j = 0
-                    while detections[0, i, j, 0] > conf_th:
-                        score = detections[0, i, j, 0]
-                        pt = (detections[0, i, j, 1:] * scale).cpu().numpy()
-                        bbox = (pt[0], pt[1], pt[2], pt[3], score)
-                        bboxes = np.vstack((bboxes, bbox))
-                        j += 1
-        
-        return bboxes