hthuwal / sign-language-gesture-recognition

Sign Language Gesture Recognition From Video Sequences Using RNN And CNN

sign-language-gesture-recognition-from-video-sequences

SIGN LANGUAGE GESTURE RECOGNITION FROM VIDEO SEQUENCES USING RNN AND CNN

The Paper on this work is published here

Please do cite it if you find this project useful. :)

UPDATE:

• Cleaner and understandable code.
• Replaced all manual editing with command line arguments.
• Removed Bugs due to changes in names of the operations in the inception model.
• Code Tested on a dummy dataset of three classes on google colab.

DataSet Used

• Argentinian Sign Language Gestures. The dataset is made available strictly for academic purposes by the owners. Please read the license terms carefully and cite their paper if you plan to use the dataset.

Requirements

• Install opencv. Note: pip install opencv-python does not have video capabilities. So I recommend to build it from source as described above.
• Install tensorflow:

  pip install tensorflow

• Install tflearn

  pip install tflearn

Training and Testing

1. Data Folder

Create two folders with any name say train_videos and test_videos in the project root directory. It should contain folders corresponding to each cateogry, each folder containing corresponding videos.

For example:

train_videos
├── Accept
│   ├── 050_003_001.mp4
│   ├── 050_003_002.mp4
│   ├── 050_003_003.mp4
│   └── 050_003_004.mp4
├── Appear
│   ├── 053_003_001.mp4
│   ├── 053_003_002.mp4
│   ├── 053_003_003.mp4
│   └── 053_003_004.mp4
├── Argentina
│   ├── 024_003_001.mp4
│   ├── 024_003_002.mp4
│   ├── 024_003_003.mp4
│   └── 024_003_004.mp4
└── Away
    ├── 013_003_001.mp4
    ├── 013_003_002.mp4
    ├── 013_003_003.mp4
    └── 013_003_004.mp4

2. Extracting frames

Command

• usage:

  video-to-frame.py [-h] gesture_folder target_folder
  

  Extract frames from gesture videos.

• positional arguments:

  gesture_folder:  Path to folder containing folders of videos of different
                    gestures.
  target_folder:   Path to folder where extracted frames should be kept.
  

• optional arguments:

  -h, --help      show the help message and exit
  

The code involves some hand segmentation (based on the data we used) for each frame. (You can remove that code if you are working on some other data set)

Extracting frames form training videos

python3 "video-to-frame.py" train_videos train_frames

Extract frames from gestures in train_videos to train_frames.

Extracting frames form test videos

python3 "video-to-frame.py" test_videos test_frames

Extract frames from gestures in test_videos to test_frames.

3. Retrain the Inception v3 model.

• Download retrain.py.

  curl -LO https://github.com/tensorflow/hub/raw/master/examples/image_retraining/retrain.py

  Note: This link may change in the future. Please refer Tensorflow retrain tutorial

• Run the following command to retrain the inception model.

  python3 retrain.py --bottleneck_dir=bottlenecks --summaries_dir=training_summaries/long --output_graph=retrained_graph.pb --output_labels=retrained_labels.txt --image_dir=train_frames

This will create two file retrained_labels.txt and retrained_graph.pb

For more information about the above command refer here.

4. Intermediate Representation of Videos

Command

• usage:

  predict_spatial.py [-h] [--input_layer INPUT_LAYER]    
                     [--output_layer OUTPUT_LAYER] [--test]    
                     [--batch_size BATCH_SIZE]    
                     graph frames_folder
  

• positional arguments:

  - graph                 graph/model to be executed
  - frames_folder         Path to folder containing folders of frames of
                          different gestures.
  

• optional arguments:

    -h, --help            show this help message and exit
    --input_layer INPUT_LAYER
                          name of input layer
    --output_layer OUTPUT_LAYER
                          name of output layer
    --test                passed if frames_folder belongs to test_data
    --batch_size BATCH_SIZE
                          batch Size
  

Approach 1

• Each Video is represented by a sequence of n dimensional vectors (probability distribution or output of softmax) one for each frame. Here n is the number of classes.

  On Training Data

  python3 predict_spatial.py retrained_graph.pb train_frames --batch=100

  This will create a file predicted-frames-final_result-train.pkl that will be used by RNN.

  On Test Data

  python3 predict_spatial.py retrained_graph.pb test_frames --batch=100 --test

  This will create a file predicted-frames-final_result-test.pkl that will be used by RNN.

Approach 2

• Each Video represented by a sequence of 2048 dimensional vectors (output of last Pool Layer) one for each frame

  On Training Data

  python3 predict_spatial.py retrained_graph.pb train_frames \
  --output_layer="module_apply_default/InceptionV3/Logits/GlobalPool" \
  --batch=100

  This will create a file predicted-frames-GlobalPool-train.pkl that will be used by RNN.

  On Test Data

  python3 predict_spatial.py retrained_graph.pb train_frames \
          --output_layer="module_apply_default/InceptionV3/Logits/GlobalPool" \
          --batch=100 \
          --test

  This will create a file predicted-frames-GlobalPool-test.pkl that will be used by RNN.

5. Train the RNN.

Command

• usage

  rnn_train.py [-h] [--label_file LABEL_FILE] [--batch_size BATCH_SIZE]
                  input_file_dump model_file
  

• positional arguments

  input_file_dump       file containing intermediate representation of gestures from inception model
  model_file            Name of the model file to be dumped. Model file is
                        created inside a checkpoints folder
  

• optional arguments

  -h, --help            show this help message and exit
  --label_file LABEL_FILE
                        path to label file generated by inception, default='retrained_labels.txt'
  --batch_size BATCH_SIZE
                        batch Size, default=32
  

Approach 1

python3 rnn_train.py predicted-frames-final_result-train.pkl non_pool.model

This will train the RNN model on the softmax based representation of gestures for 10 epochs and save the model with name non_pool.model in a folder named checkpoints.

Approach 2

python3 rnn_train.py predicted-frames-GlobalPool-train.pkl pool.model

This will train the RNN model on the pool layer based representation of gestures for 10 epochs and save the model with name pool.model in a folder named checkpoints.

6. Test the RNN Model

Command

• usage

  rnn_eval.py [-h] [--label_file LABEL_FILE] [--batch_size BATCH_SIZE]
                  input_file_dump model_file
  

• positional arguments

  input_file_dump       file containing intermediate representation of gestures from inception model
  model_file            Name of the model file to be used for prediction.
  

• optional arguments

  -h, --help            show this help message and exit
  --label_file LABEL_FILE
                        path to label file generated by inception, default='retrained_labels.txt'
  --batch_size BATCH_SIZE
                        batch Size, default=32
  

Approach 1

python3 rnn_eval.py predicted-frames-final_result-test.pkl non_pool.model

This will use the non_pool.model to predict the labels of the softmax based representation of the test videos. Predictions and corresponding gold labels for each test video will be dumped in to results.txt

Approach 2

python3 rnn_eval.py predicted-frames-GlobalPool-test.pkl pool.model

This will use the pool.model to predict the labels of the pool layer based representation of the test videos. Predictions and corresponding gold labels for each test video will be dumped in to results.txt

Happy Coding :)