Small NMS correction in OCR.

README.md

@@ -12,7 +12,7 @@ In order to easily run the code, you must have installed the Keras framework wit
 $ cd darknet && make
 ```
 
-**The current version was tested in an Ubuntu 16.04 machine, with Keras 2.0.6, TensorFlow 1.5.0 and Python 2.7.**
+**The current version was tested in an Ubuntu 16.04 machine, with Keras 2.2.4, TensorFlow 1.5.0 and Python 2.7.**
 
 ## Download Models
 

darknet/python/darknet.py

@@ -138,9 +138,10 @@ def detect(net, meta, image, thresh=.5, hier_thresh=.5, nms=.45):
  b = dets[j].bbox
  res.append((meta.names[i], dets[j].prob[i], (b.x, b.y, b.w, b.h)))
  res = sorted(res, key=lambda x: -x[1])
+ wh = (im.w,im.h)
  free_image(im)
  free_detections(dets, num)
- return res
+ return res,wh
 
 if __name__ == "__main__":
  #net = load_net("cfg/densenet201.cfg", "/home/pjreddie/trained/densenet201.weights", 0)

license-plate-ocr.py

@@ -7,6 +7,8 @@
 from os.path import splitext, basename
 from glob import glob
 from darknet.python.darknet import detect
+from src.label import dknet_label_conversion
+from src.utils import nms
 
 
 if __name__ == '__main__':
@@ -23,7 +25,7 @@
  ocr_net = dn.load_net(ocr_netcfg, ocr_weights, 0)
  ocr_meta = dn.load_meta(ocr_dataset)
 
- imgs_paths = glob('%s/*lp.png' % output_dir)
+ imgs_paths = sorted(glob('%s/*lp.png' % output_dir))
 
  print 'Performing OCR...'
 
@@ -33,12 +35,15 @@
 
  bname = basename(splitext(img_path)[0])
 
- R = detect(ocr_net, ocr_meta, img_path ,thresh=ocr_threshold)
+ R,(width,height) = detect(ocr_net, ocr_meta, img_path ,thresh=ocr_threshold, nms=None)
 
  if len(R):
 
- R.sort(key=lambda x: x[2][0])
- lp_str = ''.join([r[0] for r in R])
+ L = dknet_label_conversion(R,width,height)
+ L = nms(L,.45)
+
+ L.sort(key=lambda x: x.tl()[0])
+ lp_str = ''.join([chr(l.cl()) for l in L])
 
  with open('%s/%s_str.txt' % (output_dir,bname),'w') as f:
  f.write(lp_str + '\n')

src/label.py

@@ -85,6 +85,15 @@ def lwrite(file_path,labels,write_probs=True):
  fd.write('%d %f %f %f %f\n' % (cl,cc[0],cc[1],wh[0],wh[1]))
 
 
+def dknet_label_conversion(R,img_width,img_height):
+ WH = np.array([img_width,img_height],dtype=float)
+ L = []
+ for r in R:
+ center = np.array(r[2][:2])/WH
+ wh2 = (np.array(r[2][2:])/WH)*.5
+ L.append(Label(ord(r[0]),tl=center-wh2,br=center+wh2,prob=r[1]))
+ return L
+
 
 class Shape():
 

vehicle-detection.py

@@ -39,7 +39,7 @@
 
  bname = basename(splitext(img_path)[0])
 
- R = detect(vehicle_net, vehicle_meta, img_path ,thresh=vehicle_threshold)
+ R,_ = detect(vehicle_net, vehicle_meta, img_path ,thresh=vehicle_threshold)
 
  R = [r for r in R if r[0] in ['car','bus']]