updates

README.md

@@ -57,8 +57,6 @@ Pretrained models can be downloaded from [here](https://www.dropbox.com/sh/16jv2
 All materials (annotations and pretrained models) are also available in Baiduyun Disk:
 [here](https://pan.baidu.com/s/1yaOYqzcEx96z9gAkOhMnvQ) with password 95mm
 
-***NOTE:*** With some extra tricks, YOWO achieves 87.5% and 76.7% frame_mAP for UCF101-24 and J-HMDB-21 datasets, respectively.
-
 ## Running the code
 
 * Example training bash scripts are provided in 'run_ucf101-24.sh' and 'run_jhmdb-21.sh'.
@@ -99,6 +97,11 @@ python evaluation/Object-Detection-Metrics/pascalvoc.py --gtfolder PATH-TO-GROUN
 * For video_mAP, 'run_video_mAP_ucf.sh' and 'run_video_mAP_jhmdb.sh' bash scripts can be used.
 
 
+***UPDATEs:*** 
+* We have found a bug in our evaluation for calculating frame-mAP on UCF101-24 dataset (video-mAP results are same as before). We have fixed it, but the frame-mAP results for UCF101-24 are lower than before (if LFB are not used).
+* We have used freezing both 2D-CNN and 3D-CNN backbones for all models at the trainings of J-HMDB-21 dataset. This improved the performence considerable, especially for models having resource efficient 3D-CNN backbones.
+* We have implemented [Long-Term Feature Bank (LFB)](https://arxiv.org/pdf/1812.05038.pdf). Details can be found in the paper. It brings considerable improvement to UCF101-24 dataset and marginal improvement to J-HMDB-21 dataset. YOWO+LBF achieves 87.3% and 75.7% frame_mAP for UCF101-24 and J-HMDB-21 datasets, respectively.
+
 ### Citation
 If you use this code or pre-trained models, please cite the following:
 

evaluation/Object-Detection-Metrics/pascalvoc.py

@@ -207,15 +207,15 @@ def getBoundingBoxes(directory,
     '-gtformat',
     dest='gtFormat',
     metavar='',
-    default='xywh',
+    default='xyrb',
     help='format of the coordinates of the ground truth bounding boxes: '
     '(\'xywh\': <left> <top> <width> <height>)'
     ' or (\'xyrb\': <left> <top> <right> <bottom>)')
 parser.add_argument(
     '-detformat',
     dest='detFormat',
     metavar='',
-    default='xywh',
+    default='xyrb',
     help='format of the coordinates of the detected bounding boxes '
     '(\'xywh\': <left> <top> <width> <height>) '
     'or (\'xyrb\': <left> <top> <right> <bottom>)')

train.py

@@ -103,7 +103,7 @@
     print("===================================================================")
     print('loading checkpoint {}'.format(opt.resume_path))
     checkpoint = torch.load(opt.resume_path)
-    opt.begin_epoch = checkpoint['epoch']
+    opt.begin_epoch = checkpoint['epoch'] + 1
     best_fscore = checkpoint['fscore']
     model.load_state_dict(checkpoint['state_dict'])
     optimizer.load_state_dict(checkpoint['optimizer'])
@@ -241,14 +241,14 @@ def truths_length(truths):
                     detection_path = os.path.join('ucf_detections', 'detections_'+str(epoch), frame_idx[i])
                     current_dir = os.path.join('ucf_detections', 'detections_'+str(epoch))
                     if not os.path.exists('ucf_detections'):
-                        os.mkdir(current_dir)
+                        os.mkdir('ucf_detections')
                     if not os.path.exists(current_dir):
                         os.mkdir(current_dir)
                 else:
                     detection_path = os.path.join('jhmdb_detections', 'detections_'+str(epoch), frame_idx[i])
                     current_dir = os.path.join('jhmdb_detections', 'detections_'+str(epoch))
                     if not os.path.exists('jhmdb_detections'):
-                        os.mkdir(current_dir)
+                        os.mkdir('jhmdb_detections')
                     if not os.path.exists(current_dir):
                         os.mkdir(current_dir)
 

video_mAP.py

@@ -44,7 +44,7 @@
 
 
 # Create model
-model = YOWO(opt)
+model       = YOWO(opt)
 model       = model.cuda()
 model       = nn.DataParallel(model, device_ids=None) # in multi-gpu case
 print(model)