Revise remove indexes for label

Author: yukitsuji

cython_util/bbox_overlap.pyx

@@ -13,7 +13,7 @@ ctypedef np.int_t DTYPE_int_t
 def bbox_overlaps(
         np.ndarray[DTYPE_t, ndim=4] anchors,
         np.ndarray[DTYPE_int_t, ndim=3] is_inside,
-        np.ndarray[DTYPE_t, ndim=3] gt_boxes):
+        object gt_boxes):
     """
     Parameters
     ----------
@@ -27,7 +27,7 @@ def bbox_overlaps(
     cdef unsigned int K = anchors.shape[1]
     cdef unsigned int A = anchors.shape[2]
     cdef unsigned int G
-    cdef np.ndarray[DTYPE_t, ndim=4] overlaps = np.zeros((Batch_Size, K, A, 10), dtype=DTYPE)
+    cdef np.ndarray[DTYPE_t, ndim=4] overlaps = np.zeros((Batch_Size, K, A, 15), dtype=DTYPE)
     cdef np.ndarray[DTYPE_int_t, ndim=3] true_index = np.zeros((Batch_Size, K, A), dtype=DTYPE_int)
     cdef np.ndarray[DTYPE_int_t, ndim=3] false_index = np.zeros((Batch_Size, K, A), dtype=DTYPE_int)
     cdef DTYPE_t iw, ih, box_area
@@ -40,8 +40,8 @@ def bbox_overlaps(
         G = gt_boxes[b].shape[0]
         for g in range(G):
             box_area = (
-                (gt_boxes[b, g, 2] - gt_boxes[b, g, 0] + 1) *
-                (gt_boxes[b, g, 3] - gt_boxes[b, g, 1] + 1)
+                (gt_boxes[b][g, 2] - gt_boxes[b][g, 0] + 1) *
+                (gt_boxes[b][g, 3] - gt_boxes[b][g, 1] + 1)
             )
             max_overlap = 0
             max_k = 0
@@ -50,13 +50,13 @@ def bbox_overlaps(
                   for a in range(A):
                       if is_inside[b, k, a] == 1:
                           iw = (
-                              min(anchors[b, k, a, 2], gt_boxes[b, g, 2]) -
-                              max(anchors[b, k, a, 0], gt_boxes[b, g, 0]) + 1
+                              min(anchors[b, k, a, 2], gt_boxes[b][g, 2]) -
+                              max(anchors[b, k, a, 0], gt_boxes[b][g, 0]) + 1
                           )
                           if iw > 0:
                               ih = (
-                                  min(anchors[b, k, a, 3], gt_boxes[b, g, 3]) -
-                                  max(anchors[b, k, a, 1], gt_boxes[b, g, 1]) + 1
+                                  min(anchors[b, k, a, 3], gt_boxes[b][g, 3]) -
+                                  max(anchors[b, k, a, 1], gt_boxes[b][g, 1]) + 1
                               )
                               if ih > 0:
                                   ua = float(
@@ -65,7 +65,7 @@ def bbox_overlaps(
                                       box_area - iw * ih
                                   )
                                   overlaps[b, k, a, g] = iw * ih / ua
-                                  if max_overlap < ((iw * ih / ua):
+                                  if max_overlap < ((iw * ih / ua)):
                                       max_overlap = iw * ih / ua
                                       max_k = k
                                       max_a = a
@@ -79,7 +79,7 @@ def bbox_overlaps(
                       max_g = 0
                       for g in range(G):
                           if overlaps[b, k, a, g] > 0:
-                              if max_overlap < (overlaps[b, k, a, g):
+                              if max_overlap < (overlaps[b, k, a, g]):
                                   max_overlap = overlaps[b, k, a, g]
                                   max_g = g
                       if max_overlap > 0.7:
@@ -93,10 +93,10 @@ def bbox_overlaps(
                           ex_height = anchors[b, k, a, 3] - anchors[b, k, a, 1] + 1
                           ex_center_x = anchors[b, k, a, 0] + ex_width / 2.0
                           ex_center_y = anchors[b, k, a, 1] + ex_height / 2.0
-                          gt_width = gt_boxes[b, max_g, 2] - gt_boxes[b, max_g, 0] + 1
-                          gt_height = gt_boxes[b, max_g, 3] - gt_boxes[b, max_g, 1] + 1
-                          gt_center_x = gt_boxes[b, max_g, 0] + gt_width / 2.0
-                          gt_center_y = gt_boxes[b, max_g, 1] + gt_height / 2.0
+                          gt_width = gt_boxes[b][max_g, 2] - gt_boxes[b][max_g, 0] + 1
+                          gt_height = gt_boxes[b][max_g, 3] - gt_boxes[b][max_g, 1] + 1
+                          gt_center_x = gt_boxes[b][max_g, 0] + gt_width / 2.0
+                          gt_center_y = gt_boxes[b][max_g, 1] + gt_height / 2.0
 
                           anchors[b, k, a, 0] = (gt_center_x - ex_center_x) / (ex_width)
                           anchors[b, k, a, 1] = (gt_center_y - ex_center_y) / (ex_height)

rpn/rpn.py

@@ -1,12 +1,15 @@
 # -*- coding: utf-8 -*-
 
 import sys
+sys.path.append("/home/katou01/code/Faster_RCNN_tensorflow")
+sys.path.append("/home/katou01/code/Faster_RCNN_tensorflow/util")
+sys.path.append("/home/katou01/code/Faster_RCNN_tensorflow/cython_util")
 import glob
 import cv2
 import numpy as np
 # from vgg16 import vgg16
 from input_kitti import *
-from util import *
+from data_util import *
 from parse_xml import parseXML
 from base_vgg16 import Vgg16
 import tensorflow as tf
@@ -16,7 +19,7 @@
 from bool_anchors_inside_image import batch_inside_image
 from generate_anchors import generate_anchors
 # from utility.image.data_augmentation.flip import Flip
-sys.path.append("/Users/tsujiyuuki/env_python/code/my_code/Data_Augmentation")
+# sys.path.append("/Users/tsujiyuuki/env_python/code/my_code/Data_Augmentation")
 
 """
 ・collect dataset of cars
@@ -164,9 +167,9 @@ def rpn_loss(rpn_cls, rpn_bbox, g_bbox_regression, true_index, false_index):
     return total_loss, cls_loss, bbox_loss
 
 
-def create_Labels_For_Loss(gt_boxes, feat_stride=16, feature_shape=(64, 64), \
+def create_Labels_For_Loss(gt_boxes, feat_stride=16, feature_shape=(64, 19), \
                            scales=np.array([8, 16, 32]), ratios=[0.5, 0.8, 1], \
-                           image_size=(500, 1000)):
+                           image_size=(300, 1000)):
     """This Function is processed before network input
     Number of Candicate Anchors is Feature Map width * heights
     Number of Predicted Anchors is Batch Num * Feature Map Width * Heights * 9
@@ -180,8 +183,8 @@ def create_Labels_For_Loss(gt_boxes, feat_stride=16, feature_shape=(64, 64), \
 candicate_anchors, true_index, false_index = create_Labels_For_Loss(gt_boxes)
 candicate_anchors[true_index==1][97]
     """
-    # import time
-    # func_start = time.time()
+    import time
+    func_start = time.time()
     width = feature_shape[0]
     height = feature_shape[1]
     batch_size = gt_boxes.shape[0]
@@ -201,25 +204,35 @@ def create_Labels_For_Loss(gt_boxes, feat_stride=16, feature_shape=(64, 64), \
     candicate_anchors = centers.reshape(batch_size, K, 1, 4) + anchors # [Batch, K, A, 4]
 
     # shape is [B, K, A]
-    is_inside = batch_inside_image(candicate_anchors, image_size[0], image_size[1])
+    is_inside = batch_inside_image(candicate_anchors, image_size[1], image_size[0])
 
     # candicate_anchors: Shape is [Batch, K, A, 4]
     # gt_boxes: Shape is [Batch, G, 4]
     # true_index: Shape is [Batch, K, A]
     # false_index: Shape is [Batch, K, A]
+    # print gt_boxes.shape
+    # print gt_boxes
+    # print np.array(gt_boxes,dtype=np.float64).dtype
+    # print gt_boxes.dtype
+    print gt_boxes[0]
+#     box_area = (
+# (gt_boxes[b, g, 2] - gt_boxes[b, g, 0] + 1) *
+# (gt_boxes[b, g, 3] - gt_boxes[b, g, 1] + 1)
+# )
     candicate_anchors, true_index, false_index = bbox_overlaps(
         np.ascontiguousarray(candicate_anchors, dtype=np.float),
         is_inside,
-        np.ascontiguousarray(gt_boxes, dtype=np.float))
+        gt_boxes)
 
     for i in range(batch_size):
         true_where = np.where(true_index[i] == 1)
         num_true = len(true_where[0])
         if num_true > 64:
             select = np.random.choice(num_true, num_true - 64, replace=False)
             num_true = 64
-            true_where = remove_extraboxes(true_where[0], true_where[1], select)
-            true_index[i, true_where] = 0
+            batch = np.ones((select.shape[0]), dtype=np.int) * i
+            true_where = remove_extraboxes(true_where[0], true_where[1], select, batch)
+            true_index[true_where] = 0
 
         false_where = np.where(false_index[i] == 1)
         num_false = len(false_where[0])
@@ -228,7 +241,7 @@ def create_Labels_For_Loss(gt_boxes, feat_stride=16, feature_shape=(64, 64), \
         false_where = remove_extraboxes(false_where[0], false_where[1], select, batch)
         false_index[false_where] = 0
 
-    # print "time", time.time() - func_start
+    print "time", time.time() - func_start
     return candicate_anchors, true_index, false_index
 
 def create_ROIs_For_RCNN(model, g_labels, feat_stride=16):
@@ -291,11 +304,24 @@ def train_rpn(batch_size, image_dir, label_dir, epoch=101, label_type="txt", lr=
 
     image_dir = "/home/katou01/download/training/image_2/*.png"
     label_dir = "/home/katou01/download/training/label_2/*.txt"
+    import time
+    start = time.time()
     images, labels = get_ALL_Image(image_dir, label_dir)
-    create_Labels_For_Loss(gt_boxes, feat_stride=16, feature_shape=(64, 64), \
-                               scales=np.array([8, 16, 32]), ratios=[0.5, 0.8, 1], \
-                               image_size=(500, 1000)):
-    print images.shape, labels.shape
+    print "labels"
+    print labels.shape
+    print "images"
+    print images.shape
+    candicate_anchors, true_index, false_index = create_Labels_For_Loss(labels, feat_stride=16, feature_shape=(64, 19), \
+                               scales=np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32]), ratios=[0.1, 0.2, 0.3, 0.5, 0.8, 1, 1.2], \
+                               image_size=(302, 1000))
+    print time.time() - start
+    print images[0].shape, labels.shape
+
+    print true_index[true_index==1].shape
+    print false_index[false_index==1].shape
+
+    print true_index[0, true_index[0]==1].shape
+    print false_index[false_index==1].shape
     #
     # image = im.open("./test_images/test1.jpg")
     # image = np.array(image, dtype=np.float32)

util/util.py renamed to util/data_util.py

@@ -5,7 +5,7 @@
 import cv2
 import numpy as np
 # from vgg16 import vgg16
-from input_kitti import *
+# from input_kitti import *
 from parse_xml import parseXML
 from bbox_transform import *
 from base_vgg16 import Vgg16

util/input_kitti.py

@@ -8,7 +8,7 @@
 import glob
 import math
 from parse_xml import parseXML
-from util import *
+from data_util import *
 import matplotlib.pyplot as plt
 
 def read_label_from_txt(label_path):
@@ -86,6 +86,8 @@ def get_ALL_Image(image_dir, label_dir):
         images     (list): ndarray Images of datasets.
         pred_bboxes(ndarray): rescaled bbox Label. Shapeis [Batch, ?, 4](x, y, w, h)
     """
+    import time
+    start = time.time()
     # 車が含まれている画像のみラベルと一緒に読み込む
     image_pathlist = 0 #load_for_detection(label_dir)
     dataset_img_list = [] # len(dataset_img_list) == Number of Datasets Images
@@ -98,7 +100,9 @@ def get_ALL_Image(image_dir, label_dir):
     label_pathlist.sort()
 
     for index, (image_path, label_path) in enumerate(zip(image_pathlist, label_pathlist)):
-        if index == 100:
+        if index < 64:
+            continue
+        if index == 128:
             break
         img = cv2.imread(image_path)
         label = read_label_from_txt(label_path)
@@ -108,4 +112,5 @@ def get_ALL_Image(image_dir, label_dir):
         dataset_img_list.append(img)
         g_bboxes.append(label)
 
+    print time.time() - start
     return np.array(dataset_img_list), np.array(g_bboxes)