amirziai/kissing-detector

Kissing Detector - CS231n project


repo name	amirziai/kissing-detector
repo link	https://github.com/amirziai/kissing-detector
homepage
language	Jupyter Notebook
size (curr.)	17182 kB
stars (curr.)	50
created	2019-05-16
license	MIT License

Kissing Detector

Detect kissing scenes in a movie using both audio and video features.

Project for Stanford CS231N

Resources

Running the code

Use Python 3.6+

python3 experiments.py

this will run the experiments in params.py specified by the experiments dictionary.

Requirements

This is a PyTorch project. Look at requirements.txt for more details.

Build dataset

The following will build the dataset for training. You need to provide path to video segments.

from pipeline import BuildDataset

videos_and_labels = [
    # (file name in base_path, label) where label is 1 for kissing and 0 for not kissing
    ('movies_casino_royale_2006_kissing_1.mp4', 1),
    ('movies_casino_royale_2006_kissing_2.mp4', 1),
    ('movies_casino_royale_2006_kissing_3.mp4', 1),
    ('movies_casino_royale_2006_not_1.mp4', 0),
    ('movies_casino_royale_2006_not_2.mp4', 0),
    ('movies_casino_royale_2006_not_3.mp4', 0),
    
    ('movies_goldeneye_1995_kissing_1.mp4', 1),
    ('movies_goldeneye_1995_kissing_2.mp4', 1),
    ('movies_goldeneye_1995_kissing_3.mp4', 1),
    ('movies_goldeneye_1995_not_1.mp4', 0),
    ('movies_goldeneye_1995_not_2.mp4', 0),
    ('movies_goldeneye_1995_not_3.mp4', 0),
]

builder = BuildDataset(base_path='path/to/movies',
                 videos_and_labels=videos_and_labels,
                 output_path='/path/to/output',
                 test_size=1 / 3)  # set aside 1 / 3 of data for validation
builder.build_dataset()

Detect kissing segments in a given video

from segmentor import Segmentor
import utils

# download model.pkl from https://drive.google.com/file/d/1RlvvdInTXtJikGv_ZbHcKoblCypN1Z0A/view?usp=sharing
# or train your own
model = utils.unpickle('model.pkl')  # pickled PyTorch model 
s = Segmentor(model, min_frames=10, threshold=0.7)

# For YouTube clip Hot Summer Nights - Kiss Scene (Maika Monroe and Timothee Chalamet)
# at https://www.youtube.com/watch?v=GG5HmLQ_Fx0
# v=XXX is the YouTube ID, pass that here 
s.visualize_segments_youtube('GG5HmLQ_Fx0')

# alternatively you can provide a path to a local mp4 file
s.visualize_segments('path/to/file.mp4')

See examples in examples/detector.ipynb.