add detect

README.md

@@ -1,2 +1,7 @@
 # chainer-SSD
-uncompleted
+Single Shot MultiBox Detector
+chainer用
+まだ調整中（殴り書き）
+
+[元のリポジトリ](https://github.com/weiliu89/caffe/tree/ssd).
+[元の論文](http://arxiv.org/abs/1512.02325).

detect.py

@@ -1,12 +1,14 @@
 import argparse
 
 import numpy as np
-import cv2
+#import cv2
 import skimage.io
+import skimage.draw
 from skimage.transform import resize
 import chainer
 from chainer import serializers
 
+
 import ssd_net
 
 
@@ -16,6 +18,7 @@
 args = parser.parse_args()
 mean = np.array([104,117,123])
 img = skimage.img_as_float(skimage.io.imread(args.path, as_grey=False)).astype(np.float32)
+moto = img
 img = resize(img, (300,300))
 img = img*255 - mean[::-1]
 img = img.transpose(2, 0, 1)[::-1]
@@ -27,14 +30,16 @@
 
 def nms(bboxes, scores, score_th, nms_th, top_k):
     score_iter = 0
-    score_index = scores.argsort[::-1][:top_k]
+    score_index = scores.argsort()[::-1][:top_k]
     indices = []
     while(score_iter < len(score_index)):
         idx = score_index[score_iter]
         keep = True
-        for i in range(indices):
+        print(idx)
+        for i in range(len(indices)):
             if keep:
                 kept_idx = indices[i]
+                print(bboxes[idx], bboxes[kept_idx])
                 overlap = IoU(bboxes[idx], bboxes[kept_idx])
                 keep = overlap <= nms_th
             else:
@@ -43,30 +48,51 @@ def nms(bboxes, scores, score_th, nms_th, top_k):
             indices.append(idx)
         score_iter+=1
     return indices
-
+"""
+def nms(bboxes, nms_th, top_k):
+    bboxes = bboxes[:top_k]
+    indices = []
+    for bbox in bboxes:
+        keep = True
+        for i in range(indices):
+            if keep:
+                kept_idx = indices[i]
+                overlap = IoU(bbox, bboxes[kept_idx])
+                keep = overlap <= nms_th
+            else:
+                break
+        if keep:
+            indices.append(idx)
+    return indices
+"""
 def IoU(a, b):
-    U = union(a, b)
+    #U = union(a, b)
     I = intersection(a, b)
     if not I:
         return 0
     a_ = (a[2]-a[0])*(a[3]-a[1])
     b_ = (b[2]-b[0])*(b[3]-b[1])
+    if a_ <=0 or b_ <= 0:
+        return 1
     i = (I[2]-I[0])*(I[3]-I[1])
-    return a_ + b_ - i*2
-
-def union(a,b):
-  x1 = min(a[0], b[0])
-  y1 = min(a[1], b[1])
-  w = max(a[0]+a[2], b[0]+b[2]) - x
-  h = max(a[1]+a[3], b[1]+b[3]) - y
-  return (x, y, w, h)
+    return i/(a_ + b_ - i*2)
 
 def intersection(a,b):
   x1 = max(a[0], b[0])
   y1 = max(a[1], b[1])
   x2 = min(a[2], b[2])
-  y1 = min(a[1], b[1])
+  y2 = min(a[3], b[3])
   w = x2 - x1
   h = y2 - y1
   if w<0 or h<0: return () # or (0,0,0,0) ?
-  return (x, y, w, h)
+  return (x1, y1, x2, y2)
+
+a=model.detection()
+for i in a:
+    conf, x1, y1, x2, y2 = i
+    x1 *= moto.shape[1]
+    x2 *= moto.shape[1]
+    y1 *=moto.shape[0]
+    y2 *=moto.shape[0]
+    rr,cc = skimage.draw.polygon_perimeter([y1, y2, y2, y1],[x1, x1, x2, x2], shape=moto.shape, clip=True)
+    moto[rr, cc]= 0

ssd_net.py

@@ -60,7 +60,23 @@ def __init__(self):
 
         )
         self.train = False
-        self.conv4_3_norm_priorbox = self.prior((38, 38), 30., 0, [2], 1, 1,(0.1, 0.1, 0.2, 0.2))
+        self.conv4_3_norm_mbox_priorbox = self.prior((38, 38), 30., 0, [2], 1, 1,(0.1, 0.1, 0.2, 0.2))
+        self.fc7_mbox_priorbox = self.prior((19, 19), 60., 114., [2, 3], 1, 1,(0.1, 0.1, 0.2, 0.2))
+        self.conv6_2_mbox_priorbox = self.prior((10, 10), 114., 168., [2, 3], 1, 1,(0.1, 0.1, 0.2, 0.2))
+        self.conv7_2_mbox_priorbox = self.prior((5, 5),168., 222., [2, 3], 1, 1,(0.1, 0.1, 0.2, 0.2))
+        self.conv8_2_mbox_priorbox = self.prior((3, 3), 222., 276., [2, 3], 1, 1,(0.1, 0.1, 0.2, 0.2))
+        self.pool6_mbox_priorbox = self.prior((1, 1), 276., 330., [2, 3], 1, 1,(0.1, 0.1, 0.2, 0.2))
+        self.mbox_prior = np.hstack([self.conv4_3_norm_mbox_priorbox,
+                                     self.fc7_mbox_priorbox,
+                                     self.conv6_2_mbox_priorbox,
+                                     self.conv7_2_mbox_priorbox,
+                                     self.conv8_2_mbox_priorbox,
+                                     self.pool6_mbox_priorbox])
+
+
+
+
+
 
     def __call__(self, x, t):
 
@@ -166,11 +182,12 @@ def __call__(self, x, t):
                                    self.h_conv8_2_mbox_conf_flat,
                                    self.h_pool6_mbox_conf_flat],axis=0)
 
-        self.mbox_reahpe = F.reshape(self.mbox_conf,(1,7308,21))
-        self.mbox_reahpe_softmax = F.softmax(self.mbox_reahpe)
-        self.loss = self.loss_func(h, t)
-        self.accuracy = self.loss
-        return self.loss
+        self.mbox_conf_reahpe = F.reshape(self.mbox_conf,(7308,21))
+        self.mbox_conf_softmax = F.softmax(self.mbox_conf_reahpe)
+        if self.train:
+            self.loss = self.loss_func(h, t)
+            self.accuracy = self.loss
+            return self.loss
 
     def prior(self, h, min_size, max_size,aspect, flip, clip, variance):
         aspect_ratio = [1.]
@@ -181,7 +198,7 @@ def prior(self, h, min_size, max_size,aspect, flip, clip, variance):
         img_width = img_height = 300.
         step_x = img_width / float(width)
         step_y = img_width / float(height)
-        top_data=np.zeros(height * width * len(aspect_ratio) * 4)
+        top_data=np.zeros(height * width * (len(aspect_ratio) + bool(max_size))* 4 )
         idx=0
         for h in range(height):
             for w in range(width):
@@ -228,7 +245,7 @@ def prior(self, h, min_size, max_size,aspect, flip, clip, variance):
                     top_data[i] = 1
                 elif top_data[i] < 0:
                     top_data[i] = 0
-        val_data=np.zeros(height * width * len(aspect_ratio) * 4)
+        val_data=np.zeros(height * width * (len(aspect_ratio) + bool(max_size))* 4 )
         if len(variance)==1:
             pass
         else:
@@ -241,7 +258,23 @@ def prior(self, h, min_size, max_size,aspect, flip, clip, variance):
                             count+=1
         return np.vstack([top_data, val_data])
 
-    def decoder(prior, loc, prior_data):
+    def detection(self):
+        prior = np.reshape(self.mbox_prior,(2, 7308, 4))
+        loc = np.reshape(self.mbox_loc.data, (7308, 4))
+        conf = self.mbox_conf_softmax.data
+        cand = []
+        for label in range(1,21):
+            l = conf[:,label].argsort()
+            label_cand = np.array([np.hstack([conf[i, label] ,self.decoder(prior[0, i], loc[i], prior[1, i])]) for i in l if conf[i,label] > 0.1])
+            if label_cand.any():
+                k = self.nms(label_cand[:,1:], label_cand[:,0], 0.1, 0.45, 200)
+                for i in k:
+                    cand.append(label_cand[i])
+        cand = np.array(cand)
+        cand = cand[np.where(cand[:,0]>=0.6)]
+        return cand
+
+    def decoder(self, prior, loc, prior_data):
         bbox_data = np.array([0]*4,dtype=np.float32)
         p_xmin, p_ymin, p_xmax, p_ymax= prior
         xmin, ymin, xmax, ymax= loc
@@ -254,7 +287,53 @@ def decoder(prior, loc, prior_data):
         decode_bbox_width = np.exp(prior_data[2] * xmax) * prior_width;
         decode_bbox_height = np.exp(prior_data[3] * ymax) * prior_height;
         bbox_data[0] = decode_bbox_center_x - decode_bbox_width / 2.
-        bbox_data[1] = decode_bbox_center_y - decode_bbox_height / 2.;
-        bbox_data[2] = decode_bbox_center_x + decode_bbox_width / 2.;
-        bbox_data[3] = decode_bbox_center_y + decode_bbox_height / 2.;
+        bbox_data[1] = decode_bbox_center_y - decode_bbox_height / 2.
+        bbox_data[2] = decode_bbox_center_x + decode_bbox_width / 2.
+        bbox_data[3] = decode_bbox_center_y + decode_bbox_height / 2.
         return bbox_data
+
+    def nms(self, bboxes, scores, score_th, nms_th, top_k):
+        score_iter = 0
+        score_index = scores.argsort()[::-1][:top_k]
+        indices = []
+        while(score_iter < len(score_index)):
+            idx = score_index[score_iter]
+            keep = True
+            cand_bbox=bboxes[idx]
+            if cand_bbox[0] == cand_bbox[2] or cand_bbox[1] == cand_bbox[3]:
+                keep=False
+            print(idx)
+            for i in range(len(indices)):
+                if keep:
+                    kept_idx = indices[i]
+                    overlap = self.IoU(bboxes[idx], bboxes[kept_idx])
+                    print("over = ",overlap)
+                    keep = overlap <= nms_th
+                else:
+                    break
+            if keep:
+                indices.append(idx)
+            score_iter+=1
+        return indices
+
+    def IoU(self, a, b):
+        I = self.intersection(a, b)
+        if not I:
+            return 0
+        a_ = (a[2]-a[0])*(a[3]-a[1])
+        b_ = (b[2]-b[0])*(b[3]-b[1])
+        print(a_, b_)
+        if a_ <=0 or b_ <= 0:
+            return 1
+        i = (I[2]-I[0])*(I[3]-I[1])
+        return i/(a_ + b_ - i*2)
+
+    def intersection(self, a,b):
+        x1 = max(a[0], b[0])
+        y1 = max(a[1], b[1])
+        x2 = min(a[2], b[2])
+        y2 = min(a[3], b[3])
+        w = x2 - x1
+        h = y2 - y1
+        if w<0 or h<0: return () # or (0,0,0,0) ?
+        return (x1, y1, x2, y2)