add ctw1500

README.md

@@ -7,9 +7,6 @@
 * Polygon2
 * OpenCV 3+ (for c++ version pse)
 
-# Todo
-* CTW1500 train and test
-
 ## Introduction
 Progressive Scale Expansion Network (PSENet) is a text detector which is able to well detect the arbitrary-shape text in natural scene.
 
@@ -32,6 +29,13 @@ CUDA_VISIBLE_DEVICES=0 python test_ic15.py --scale 1 --resume [path of model]
 | PSENet-1s (ResNet50) | pretrain on IC17 MLT | 86.92 | 84.5 | 85.69 | 3.8 | [baiduyun](https://pan.baidu.com/s/1oKVxHKuT3hdzDUmksbcgAQ)(extract code: aieo) |
 | PSENet-4s (ResNet50) | pretrain on IC17 MLT | 86.1 | 83.77 | 84.92 | 3.8 | [baiduyun](https://pan.baidu.com/s/1oKVxHKuT3hdzDUmksbcgAQ)(extract code: aieo) |
 
+### [SCUT-CTW1500](https://github.com/Yuliang-Liu/Curve-Text-Detector)
+| Method | Extra Data | Precision (%) | Recall (%) | F-measure (%) | FPS | Model |
+| - | - | - | - | - | - | - |
+| PSENet-1s (ResNet50) | - | 80.57 | 75.55 | 78.0 | 3.9 | [baiduyun](https://pan.baidu.com/s/1BqJspFwBmHjoqlE0jOrJQg)(extract code: ksv7) |
+| PSENet-1s (ResNet50) | pretrain on IC17 MLT | 84.84| 79.73 | 82.2 | 3.9 | todo |
+| PSENet-4s (ResNet50) | pretrain on IC17 MLT | 82.09 | 77.84 | 79.9 | 8.4 | todo |
+
 ## Performance (old version paper on arxiv)
 ### [ICDAR 2015](http://rrc.cvc.uab.es/?ch=4&com=evaluation&task=1) (training with ICDAR 2017 MLT)
 | Method | Precision (%) | Recall (%) | F-measure (%) | 

dataset/__init__.py

@@ -1,2 +1,5 @@
 from dataset.icdar2015_loader import IC15Loader
-from dataset.icdar2015_test_loader import IC15TestLoader
+from dataset.icdar2015_test_loader import IC15TestLoader
+
+from dataset.ctw1500_loader import CTW1500Loader
+from dataset.ctw1500_test_loader import CTW1500TestLoader

dataset/ctw1500_loader.py

@@ -147,7 +147,6 @@ def shrink(bboxes, rate, max_shr=20):
 
 class CTW1500Loader(data.Dataset):
     def __init__(self, is_transform=False, img_size=None, kernel_num=7, min_scale=0.4):
-        self.split = split
         self.is_transform = is_transform
 
         self.img_size = img_size if (img_size is None or isinstance(img_size, tuple)) else (img_size, img_size)

dataset/ctw1500_test_loader.py

@@ -11,8 +11,8 @@
 import pyclipper
 import Polygon as plg
 
-ctw_root_dir = '/home/shared/ctw1500/'
-ctw_test_data_dir = ctw_root_dir + 'tmp/'
+ctw_root_dir = './data/CTW1500/'
+ctw_test_data_dir = ctw_root_dir + 'test/text_image/'
 
 random.seed(123456)
 
@@ -25,7 +25,7 @@ def get_img(img_path):
         raise
     return img
 
-def scale(img, long_size=2240):
+def scale(img, long_size=1280):
     h, w = img.shape[0:2]
     scale = long_size * 1.0 / max(h, w)
     img = cv2.resize(img, dsize=None, fx=scale, fy=scale)
@@ -63,10 +63,10 @@ def __getitem__(self, index):
 
         img = get_img(img_path)
 
-        scaled_img = scale(img, self.long_size)    
-        scaled_img = Image.fromarray(img)
-        scaled_img = img.convert('RGB')
-        scaled_img = transforms.ToTensor()(img)
-        scaled_img = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])(img)
-
-        return img[:, :, [2, 1, 0]]
+        scaled_img = scale(img, self.long_size)
+        scaled_img = Image.fromarray(scaled_img)
+        scaled_img = scaled_img.convert('RGB')
+        scaled_img = transforms.ToTensor()(scaled_img)
+        scaled_img = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])(scaled_img)
+        
+        return img[:, :, [2, 1, 0]], scaled_img

eval/ctw1500/eval_ctw1500.py

@@ -0,0 +1,94 @@
+import file_util
+import Polygon as plg
+import numpy as np
+
+pred_root = '../../outputs/submit_ctw1500/'
+gt_root = '../../data/CTW1500/test/text_label_curve/'
+
+def get_pred(path):
+    lines = file_util.read_file(path).split('\n')
+    bboxes = []
+    for line in lines:
+        if line == '':
+            continue
+        bbox = line.split(',')
+        if len(bbox) % 2 == 1:
+            print path
+        bbox = [(int)(x) for x in bbox]
+        bboxes.append(bbox)
+    return bboxes
+
+def get_gt(path):
+    lines = file_util.read_file(path).split('\n')
+    bboxes = []
+    for line in lines:
+        if line == '':
+            continue
+        # line = util.str.remove_all(line, '\xef\xbb\xbf')
+        # gt = util.str.split(line, ',')
+        gt = line.split(',')
+
+        x1 = np.int(gt[0])
+        y1 = np.int(gt[1])
+
+        bbox = [np.int(gt[i]) for i in range(4, 32)]
+        bbox = np.asarray(bbox) + ([x1, y1] * 14)
+
+        bboxes.append(bbox)
+    return bboxes
+
+def get_union(pD,pG):
+    areaA = pD.area();
+    areaB = pG.area();
+    return areaA + areaB - get_intersection(pD, pG);        
+
+def get_intersection(pD,pG):
+    pInt = pD & pG
+    if len(pInt) == 0:
+        return 0
+    return pInt.area()
+
+if __name__ == '__main__':
+    th = 0.5
+    pred_list = file_util.read_dir(pred_root)
+
+    tp, fp, npos = 0, 0, 0
+
+    for pred_path in pred_list:
+        preds = get_pred(pred_path)
+        gt_path = gt_root + pred_path.split('/')[-1]
+        gts = get_gt(gt_path)
+        npos += len(gts)
+
+        cover = set()
+        for pred_id, pred in enumerate(preds):
+            pred = np.array(pred)
+            pred = pred.reshape(pred.shape[0] / 2, 2)
+            # if pred.shape[0] <= 2:
+            #     continue
+            pred_p = plg.Polygon(pred)
+
+            flag = False
+            for gt_id, gt in enumerate(gts):
+                gt = np.array(gt)
+                gt = gt.reshape(gt.shape[0] / 2, 2)
+                gt_p = plg.Polygon(gt)
+
+                union = get_union(pred_p, gt_p)
+                inter = get_intersection(pred_p, gt_p)
+
+                if inter * 1.0 / union >= th:
+                    if gt_id not in cover:
+                        flag = True
+                        cover.add(gt_id)
+            if flag:
+                tp += 1.0
+            else:
+                fp += 1.0
+
+    print tp, fp, npos
+    precision = tp / (tp + fp)
+    recall = tp / npos
+    hmean = 0 if (precision + recall) == 0 else 2.0 * precision * recall / (precision + recall)
+
+    print 'p: %.4f, r: %.4f, f: %.4f'%(precision, recall, hmean)

eval/ctw1500/file_util.py

@@ -0,0 +1,33 @@
+import os
+
+def read_dir(root):
+	file_path_list = []
+	for file_path, dirs, files in os.walk(root):
+		for file in files:
+			file_path_list.append(os.path.join(file_path, file).replace('\\', '/'))
+	file_path_list.sort()
+	return file_path_list
+
+def read_file(file_path):
+	file_object = open(file_path, 'r')
+	file_content = file_object.read()
+	file_object.close()
+	return file_content
+
+def write_file(file_path, file_content):
+	if file_path.find('/') != -1:
+		father_dir = '/'.join(file_path.split('/')[0:-1])
+		if not os.path.exists(father_dir):
+			os.makedirs(father_dir)
+	file_object = open(file_path, 'w')
+	file_object.write(file_content)
+	file_object.close()
+
+
+def write_file_not_cover(file_path, file_content):
+	father_dir = '/'.join(file_path.split('/')[0:-1])
+	if not os.path.exists(father_dir):
+		os.makedirs(father_dir)
+	file_object = open(file_path, 'a')
+	file_object.write(file_content)
+	file_object.close()

eval/eval_ctw1500.sh

@@ -0,0 +1,3 @@
+cd ctw1500
+python eval_ctw1500.py
+cd ..