diff --git a/data_utils/bboxCropper/README.md b/data_utils/bboxCropper/README.md
new file mode 100644
index 0000000..911efd0
--- /dev/null
+++ b/data_utils/bboxCropper/README.md
@@ -0,0 +1,18 @@
+### Cropping tool for dataset images
+Before running **bboxCropper.py**: update **config.cfg** with your data.
+
+
+#### Inputs
+Dataset, located at **DATASET_PATH**.
+
+
+#### Outputs
+- creates folders **CROPS_FOLDER_NAME**, **FRAMES_FOLDER_NAME**, **MASKS_FOLDER_NAME** in dataset folder.
+- crops square of **CROP_SIZE*****CROP_SIZE** around bbox center on the image, with random shift, and saves it to **CROPS_FOLDER_NAME** folder.
+- cuts out bbox content from this crop, saves it to **FRAMES_FOLDER_NAME** folder.
+- creates binary mask of the size of the crop, with True pixels under bbox and all other zeros, saves it to **MASKS_FOLDER_NAME** folder.
+
+#### About masks
+- by default, mask pixels are True in places where image inpainting required, others are False. For mask inversion: set INVERT_MASKS in cfg file.
+
+![Example](https://github.com/lacmus-foundation/lacmus/blob/master/data_utils/bboxCropper/screenshot.PNG)
diff --git a/data_utils/bboxCropper/bboxCropper.py b/data_utils/bboxCropper/bboxCropper.py
new file mode 100644
index 0000000..2fe7d3e
--- /dev/null
+++ b/data_utils/bboxCropper/bboxCropper.py
@@ -0,0 +1,146 @@
+import cv2
+from pathlib import Path
+import xml.etree.ElementTree as ET 
+import numpy as np
+import os
+import random
+
+def main():
+    
+    # open CFG file, define paths and crops shapes  
+    
+    config = open('config.cfg')
+    for line in config:
+        if line.startswith('CROP_SIZE'):
+            crop_size = int(line.split('=')[1].strip().replace('\n',''))
+        if line.startswith('DATASET_PATH'):
+            dataset_path = Path(line.split('=')[1].strip().replace('\n',''))
+        if line.startswith('CROPS_FOLDER'):
+            crops_folder = line.split('=')[1].strip().replace('\n','')
+        if line.startswith('FRAMES_FOLDER'):
+            frames_folder = line.split('=')[1].strip().replace('\n','')
+        if line.startswith('MASKS_FOLDER'):
+            masks_folder = line.split('=')[1].strip().replace('\n','')
+        if line.startswith('INVERT_MASKS'):
+            invert_masks = line.split('=')[1].strip().replace('\n','')
+            if invert_masks.lower() in ['false']:
+                invert_masks = False
+            else:
+                invert_masks = True
+            
+    images_folder      =         'JPEGImages'
+    annotations_folder =         'Annotations'  
+    config.close()
+
+    print('Dataset location: ', dataset_path)
+    
+   # Create folders for outputs
+    if frames_folder not in os.listdir(dataset_path):
+        os.mkdir(Path(dataset_path, frames_folder))
+        print('Created folder: ', Path(dataset_path, frames_folder))
+    if crops_folder not in os.listdir(dataset_path):
+        os.mkdir(Path(dataset_path, crops_folder))
+        print('Created folder: ', Path(dataset_path, crops_folder))
+    if masks_folder not in os.listdir(dataset_path):
+        os.mkdir(Path(dataset_path, masks_folder))
+        print('Created folder: ', Path(dataset_path, masks_folder))
+
+    print('Processing started...')
+        
+    # parse each annotation file
+    
+    n_files = len(os.listdir(Path(dataset_path, annotations_folder)))
+    passed_files=1
+
+    for filename in os.listdir(Path(dataset_path, annotations_folder)):
+        
+        if not filename.endswith('.xml'): continue
+            
+        fullname = Path(dataset_path, annotations_folder, filename)    
+        tree = ET.parse(fullname)    
+        root = tree.getroot()    
+        bbox_num = 0
+
+        img = cv2.imread(str(Path(dataset_path, images_folder, filename[:-3]+'jpg')))
+    
+        for rec in root:
+            
+            # get source image size
+            if rec.tag == 'size': 
+                height = int(rec.findtext('height'))
+                width = int(rec.findtext('width'))
+                
+            # list all available bboxes        
+            if rec.tag == 'object': 
+                for box in rec:
+                    if box.tag=='bndbox':
+                        
+                        # get initial bbox corners
+                        ymin = int(box.findtext('ymin'))
+                        ymax = int(box.findtext('ymax'))
+                        xmin = int(box.findtext('xmin'))
+                        xmax = int(box.findtext('xmax'))
+                        
+                        # calculate necessary padding to get crop of crop_size
+                        padding_w = int((crop_size - (xmax - xmin))/2.)
+                        padding_h = int((crop_size - (ymax - ymin))/2.)
+                        
+                        # get random shift within 25% of crop_size from bbox center
+                        random_dx = int((random.random()-.5)*.5*crop_size)
+                        random_dy = int((random.random()-.5)*.5*crop_size)
+                        
+                        # calculate crop corners
+                        new_xmin = xmin - padding_w + random_dx
+                        new_xmax = xmax + padding_w + random_dx
+                        new_ymin = ymin - padding_h + random_dy
+                        new_ymax = ymax + padding_h + random_dy
+                        
+                        # do not proceed if crop is outside of image
+                        if (new_xmin<1 or new_xmax>width-1 or new_ymin<1 or new_ymax>height-1):continue
+                        
+                        dx = new_xmax - new_xmin
+                        dy = new_ymax - new_ymin
+                        
+                        # correct crop corners to get exact crop_size
+                        if dx<crop_size:
+                            if ((new_xmax+new_xmin)/2.)<(width/2.):
+                                new_xmax+=1
+                            else:
+                                new_xmin-=1
+                                
+                        if dy<crop_size:
+                            if ((new_ymax+new_ymin)/2.)<(height/2.):
+                                new_ymax+=1
+                            else:
+                                new_ymin-=1
+                       
+                        # create crop
+                        crop = img.copy()
+                        crop = crop[new_ymin:new_ymax, new_xmin:new_xmax]
+                        
+                        # create binary mask
+                        mask = np.zeros_like(img).astype(bool)
+                        mask[ymin:ymax, xmin:xmax]=True
+                        if invert_masks: mask = np.invert(mask)
+                        mask = (mask[new_ymin:new_ymax, new_xmin:new_xmax][:,:,0]*255).astype('uint8')
+                                            
+                        # cut out bbox contents from the crop to get a frame
+                        frame = img.copy()
+                        frame[ymin:ymax, xmin:xmax] = 0
+                        frame = frame[new_ymin:new_ymax, new_xmin:new_xmax]
+                        
+                        # save all this stuff
+                        cv2.imwrite(str(Path(dataset_path, masks_folder, filename[:-4]+'_'+str(bbox_num)+'_mask.png')), mask)               
+                        cv2.imwrite(str(Path(dataset_path, frames_folder, filename[:-4]+'_'+str(bbox_num)+'_crop.jpg')), frame)
+                        cv2.imwrite(str(Path(dataset_path, crops_folder, filename[:-4]+'_'+str(bbox_num)+'_pic.jpg')), crop)
+                        
+                        # goto next bbox in current file
+                        bbox_num = bbox_num + 1
+                        
+        if passed_files in range(int(n_files/10), n_files, int(n_files/10)): print(str(int(passed_files/n_files*10*10)+1)+'% done...')
+        passed_files+=1
+
+    print('Crop completed.')                   
+                            
+if __name__=='__main__':
+    main()
diff --git a/data_utils/bboxCropper/config.cfg b/data_utils/bboxCropper/config.cfg
new file mode 100644
index 0000000..3966c45
--- /dev/null
+++ b/data_utils/bboxCropper/config.cfg
@@ -0,0 +1,6 @@
+CROP_SIZE          = 512
+DATASET_PATH       = C:\TMP\test
+CROPS_FOLDER_NAME  = Crops
+FRAMES_FOLDER_NAME = Frames
+MASKS_FOLDER_NAME  = Masks
+INVERT_MASKS       = False
diff --git a/data_utils/bboxCropper/screenshot.PNG b/data_utils/bboxCropper/screenshot.PNG
new file mode 100644
index 0000000..9adff81
Binary files /dev/null and b/data_utils/bboxCropper/screenshot.PNG differ
diff --git a/docs/train-usage.md b/docs/train-usage.md
index d3cb1bc..c2f823d 100644
--- a/docs/train-usage.md
+++ b/docs/train-usage.md
@@ -6,15 +6,14 @@ This code is borrowed from Keras Implementation of this model at https://github.
 
 ## Installation
 
-1) Clone this repository.
-2) `cd RetinaNet`
-3) Ensure numpy is installed using `pip install numpy --user`
-4) In the repository, execute `pip install . --user`.
+1) Clone this repository.`
+2) Ensure numpy is installed using `pip install numpy --user`
+3) In the repository, execute `pip install . --user`.
    Note that due to inconsistencies with how `tensorflow` should be installed,
    this package does not define a dependency on `tensorflow` as it will try to install that (which at least on Arch Linux results in an incorrect installation).
    Please make sure `tensorflow` is installed as per your systems requirements.
-5) Alternatively, you can run the code directly from the cloned  repository, however you need to run `python setup.py build_ext --inplace` to compile Cython code first.
-6) Optionally, install `pycocotools` if you want to train / test on the MS COCO dataset by running `pip install --user git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI`.
+4) To run the code directly from the cloned  repository, you need to run `python setup.py build_ext --inplace` to compile Cython code first.
+5) Optionally, install `pycocotools` if you want to train / test on the MS COCO dataset by running `pip install --user git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI`.
 
 
 ## Training on custom data set (added by Priyanka Dwivedi, Georgy Perevozchikov)
@@ -55,7 +54,7 @@ Here:
 
 If your dataset is small, you can also use the `--freeze-backbone argument` to freeze the backbone layers.
 
-You can find more arguments directly in parse_args function of [train.py](../RetinaNet/keras_retinanet/bin/train.py) file.
+You can find more arguments directly in parse_args function of [train.py](../keras_retinanet/bin/train.py) file.
 
 A model trained on a LADD dataset can be found [here](https://github.com/lizaalert/lacmus/releases/tag/0.1.1).
 
@@ -113,6 +112,6 @@ Here first path is the path to the trained model and the second would be the pat
 I created two notebooks from the original code that can be used to run inference on images and on videos.
 
 The notebooks are:
-* [RetinaNet/ResNet50RetinaNet-Image.ipynb](../RetinaNet/ResNet50RetinaNet-Image.ipynb) : For inference on a batch of images
-* [RetinaNet/ResNet50RetinaNet-Video.ipynb](../RetinaNet/ResNet50RetinaNet-Video.ipynb) : For inference on a video
+* [RetinaNet/ResNet50RetinaNet-Image.ipynb](../ResNet50RetinaNet-Image.ipynb) : For inference on a batch of images
+* [RetinaNet/ResNet50RetinaNet-Video.ipynb](../ResNet50RetinaNet-Video.ipynb) : For inference on a video
 
diff --git a/keras_retinanet/bin/evaluate.py b/keras_retinanet/bin/evaluate.py
index 1a10f23..3c03aa1 100755
--- a/keras_retinanet/bin/evaluate.py
+++ b/keras_retinanet/bin/evaluate.py
@@ -28,6 +28,7 @@
 from .. import models
 from ..preprocessing.csv_generator import CSVGenerator
 from ..preprocessing.pascal_voc import PascalVocGenerator
+from ..utils.anchors import make_shapes_callback
 from ..utils.config import read_config_file, parse_anchor_parameters
 from ..utils.eval import evaluate
 from ..utils.gpu import setup_gpu
@@ -35,9 +36,13 @@
 from ..utils.tf_version import check_tf_version
 
 
-def create_generator(args):
+def create_generator(args, preprocess_image):
     """ Create generators for evaluation.
     """
+    common_args = {
+        'preprocess_image' : preprocess_image,
+    }
+
     if args.dataset_type == 'coco':
         # import here to prevent unnecessary dependency on cocoapi
         from ..preprocessing.coco import CocoGenerator
@@ -49,6 +54,7 @@ def create_generator(args):
             image_max_side=args.image_max_side,
             config=args.config,
             shuffle_groups=False,
+            **common_args
         )
     elif args.dataset_type == 'pascal':
         validation_generator = PascalVocGenerator(
@@ -58,6 +64,7 @@ def create_generator(args):
             image_max_side=args.image_max_side,
             config=args.config,
             shuffle_groups=False,
+            **common_args
         )
     elif args.dataset_type == 'csv':
         validation_generator = CSVGenerator(
@@ -67,6 +74,7 @@ def create_generator(args):
             image_max_side=args.image_max_side,
             config=args.config,
             shuffle_groups=False,
+            **common_args
         )
     else:
         raise ValueError('Invalid data type received: {}'.format(args.dataset_type))
@@ -129,7 +137,8 @@ def main(args=None):
         args.config = read_config_file(args.config)
 
     # create the generator
-    generator = create_generator(args)
+    backbone = models.backbone(args.backbone)
+    generator = create_generator(args, backbone.preprocess_image)
 
     # optionally load anchor parameters
     anchor_params = None
@@ -140,6 +149,8 @@ def main(args=None):
     print('Loading model, this may take a second...')
     model = models.load_model(args.model, backbone_name=args.backbone)
 
+    generator.compute_shapes = make_shapes_callback(model)
+
     # optionally convert the model
     if args.convert_model:
         model = models.convert_model(model, anchor_params=anchor_params)