omkarthawakar / hmdb51_data_gen

script to extract frames from HMDB51 dataset and create train, test and val split

HMDB51 data generation for training of Non Local Neural Network

Requirement

PIL
numpy
cv2
glob

HMDB51 datset folder contains following actions and each action contains different videos

----HMDB51
        |-brush_hair
                  |-April_09_brush_hair_u_nm_np1_ba_goo_0.avi
                  .....
                  |-Slave_brush_my_hair_brush_hair_u_cm_np2_le_goo_2.avi
        |-cartwheel
                  |-
                  .....
                  |-
        
        |-catch
                  |-
                  .....
                  |-
        
        |-.......
        |-wave
                  |-21_wave_u_nm_np1_fr_goo_5.avi
                  .....
                  |-
                  

To extract frames from each HMDB51 actions run following code

python frameExtract.py

To create train test split we require annotation of each video in actionset which is found in splits folder downloaded from HMDB51

Train_Test split file contains following structure

----splits
        |-brush_hair_test_split1.txt
        |-brush_hair_test_split2.txt
        |-brush_hair_test_split3.txt
        |-cartwheel_test_split1.txt
        |-.......
        |-.......
        |- 
        |-action_name_test_split{}.txt

For getting train test splits run following code

python create_train_test_split.py

Above script takes about an hour to read whole dataset and saves train,test and validation data in following files respectively.

---|-train_data.txt
   |-train_labels.txt
   |-test_data.txt
   |-test_labels.txt
   |-val_data.txt
   |-val_labels.txt

For furthur use load train,test and val data directly by running

np.loadtxt('train_data.txt')