xingsh2000/ss-videocaptioning

Semantically Sensible Video Captioning

SS-VideoCaptioning

This repository contains the Tensorflow implementation of our model "Semantically Sensible Video Captioning (SSVC)"
[Code] [Paper] [ArXiv]

Authors

Md. Mushfiqur Rahman, Thasin Abedin, Khondokar S. S. Prottoy, Ayana Moshruba, Fazlul Hasan Siddiqui

Requirements

Install the following dependencies before running the model

• Tensorflow 2.0 install
• tqdm pip install tqdm
• sklearn pip install -U scikit-learn
• nltk pip install nltk

Directory structure

-root
  -glove.6B.100d.txt
  -MSVD_captions.csv
  -models_and_utils
    -models.py
    -utils.py
  -data_picle
    -train
      -filename1.pkl
      -filename2.pkl
      ...
    -test
      -filename1.pkl
      -filename2.pkl
      ...
    -validation
      -filename1.pkl
      -filename2.pkl
      ...
    -train.csv
    -test.csv
    -validation.csv

Train and Evaluate

• Download and extract 'glove.6B.100d.txt' link
• Download the MSVD dataset and create corresponding pickle files using vid2frames.ipynb. Split the data in train-test-val sets.

  Alternate step: Download and extract 'data_pickle.zip'. This compressed file already contains the pickles files of MSVD dataset

• run the train.ipynb file

  This file has a detailed list of options. Change the options to adjust the model according to requirements

• Train and evaluation codes are inside the python notebook

Sample Outputs

SSVC: "A woman is cutting a piece of meat"
GT: "a woman is cutting into the fatty areas of a pork chop"
SS score: 1.0, BLEU1: 1.0, BLEU2: 1.0, BLEU3: 1.0, BLEU4: 1.0

SSVC: "A person is slicing tomato"
GT: "Someone wearing blue rubber gloves is slicing a tomato with a large knife"
SS score: 0.825, BLEU1: 1.0, BLEU2: 1.0, BLEU3: 1.0, BLEU4: 1.0

SSVC: "A woman is cutting a piece of meat"
GT: "a woman is cutting into the fatty areas of a pork chop"
SS score: 0.94, BLEU1: 1.0, BLEU2: 0.84, BLEU3: 0.61, BLEU4: 0.0