Fixed rectangle creation and added support for testing on N2DH dataset

toolkit/datasets/__init__.py

@@ -36,7 +36,11 @@ class DatasetFactory(object):
             dataset = TrackingNetDataset(**kwargs)
         elif 'GOT-10k' == name:
             dataset = GOT10kDataset(**kwargs)
+        # Kristyna Janku: Added option for N2DH dataset, no special attributes are needed, so I just reused
+        # the GOT10kDataset class
+        elif 'N2DH' == name:
+            dataset = GOT10kDataset(**kwargs)
         else:
-            raise Exception("unknow dataset {}".format(kwargs['name']))
+            raise Exception("unknown dataset {}".format(kwargs['name']))
         return dataset
 

training_dataset/Fluo-N2DH-SIM+/gen_json.py

@@ -4,6 +4,7 @@ import os
 import glob
 import json
 import cv2
+import numpy as np
 
 data_path = 'data'
 video_list = ['01', '02']
@@ -42,11 +43,14 @@ for video in video_list:
                 mask = cv2.convertScaleAbs(mask)
                 # Checking if tracked object really is in segmented image
                 if mask.max() == 0:
-                    print('Video: ' + str(num_videos) + ' Frame: ' + frame + ' Missing object id in seg: ' + str(obj_id))
+                    print(f'Video: {video} Frame: {frame} Missing object id in seg: {str(obj_id)}')
                 else:
                     # Get bounding box of object from contour
-                    cnt = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[0]
-                    x, y, w, h = cv2.boundingRect(cnt[0])
+                    contours = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[0]
+                    areas = [cv2.contourArea(c) for c in contours]
+                    max_index = np.argmax(areas)
+                    max_contour = contours[max_index]
+                    x, y, w, h = cv2.boundingRect(max_contour)
                     snippets[video][str(obj_id).zfill(2)][frame.zfill(6)] = [x, y, x+w, y+h]
 
     print('Video number: {:d} Snippets count: {:d}'.format(num_videos, len(snippets[video])))

training_dataset/Fluo-N2DH-SIM+/par_crop.py

@@ -99,7 +99,7 @@ def crop_video(video, crop_path, instance_size, low, high):
 
         for object_id, files in annotations[video].items():
             for filename, bbox in files.items():
-                filepath = join(data_path, video, 't' + filename[3:] + '.tif')
+                filepath = join(data_path, video, f't{filename[3:]}.tif')
 
                 # Load 16-bit TIFF image, convert it to 8-bit and perform percentile stretch
                 image_16bit = cv2.imread(filepath, cv2.IMREAD_ANYDEPTH)