By Zhi-Song Liu*, Robin Courant* and Vicky Kalogeiton

ACCV 2022 (Oral, Best Student Paper Honorable mention)

Python 3.8 OpenCV library Pytorch 1.12.0 CUDA 11.3

1. Clone code to your local computer.

git clone https://github.com/robincourant/FunnyNet.git
cd FunnyNet

2. Create working environment.

conda create --name funnynet -y python=3.8
conda activate funnynet

1. Install the dependencies.

conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

1. Run the setup script to intsall all the dependencies.

./setup.sh

1. Modify in ext/TimeSformer/timesformer/models/vit_utils.py

from torch._six import container_abcs --> import collections.abc as container_abcs

1. Comment ext/TimeSformer/timesformer/models/resnet_helper.py

from torch.nn.modules.linear import _LinearWithBias

1. Download friends data:

gdown https://drive.google.com/drive/folders/1ZM6agmEnheiyP0IIrD3Fc7DOubjyu5eO -O ./data --folder

Note: label files are strutured as follow: [season, episode, funny-label, start, end]

The dataset directory is organized as followed:

FunnyNet-data/
└── tv_show_name/
    ├── audio/
    │   ├── diff/              # `.wav` files with stereo channel difference
    │   ├── embedding/         # `.pt` files with audio embedding vectors
    │   ├── laughter/          # `.pickle` files with laughter timecodes
    │   ├── laughter_segment/  # `.wav` files with detected laughters
    │   ├── left/              # `.wav` files with the surround left channel
    │   └── raw/               # `.wav` files with extracted raw audio from videos
    ├── laughter/              # `.pk` files with laughter labels
    ├── sub/                   # `.pk` files with subtitles
    ├── episode/               # `.mkv` files with videos
    ├── audio_split/           # `.wav` files with audio 8 seconds windows
    │   ├── test_8s/
    │   ├── train_8s/
    │   └── validation_8s/
    ├── video_split/           # `.mp4` files with video 8 seconds windows
    │   ├── test_8s/
    │   ├── train_8s/
    │   └── validation_8s/
    └── sub_split/             # `.pk` files with subtitles 8 seconds windows
        ├── sub_test_8s.pk
        ├── sub_train_8s.pk
        └── sub_validation_8s.pk

Note: we cannot provide audio and video data for obvious copyright issues.

Split audio, subtitles and videos into segments of n seconds (default 8 seconds):

python data_processing/mask_audio.py DATA_DIR/audio/raw DATA_DIR/audio/laughter DATA_DIR/audio/processed
python data_processing/audio_processing.py DATA_DIR/audio/raw DATA_DIR/laughter/xx.pk DATA_DIR/audio_split
python data_processing/sub_processing.py DATA_DIR/sub DATA_DIR/laughter/xx.pk DATA_DIR/sub_split
python data_processing/video_processing.py DATA_DIR/episode DATA_DIR/laughter/xx.pk DATA_DIR/video_split

1. Train multimodality with audio and vision

python funnynet/train.py model.batch_size=BATCH_SIZE xp_name=XP_NAME data.data_dir=DATA_DIR model=avf-timesformer-byol-lstm data=avf-timesformer-byol-lstm

1. Test multimodality with audio and vision

python funnynet/evaluate.py

There is 4 scripts:

• laughter_detection/scripts/extract_audio.py: extracts from video files contained in episode/ corresponding audio tracks and saves them in audio/raw/ .

• laughter_detection/scripts/detect_laughter.py: detects laughters from audio files in audio/raw/ and saves laughter timecodes as .pickle files in audio/laughter/.

• laughter_detection/scripts/extract_laughter.py: extracts from raw audio segments in audio/raw/ each detected laughter in audio/laughter/ and saves them in audio/laughter_segment/.

• laughter_detection/scripts/evaluate_laughters.py: given directories of predicted and ground-truth laughter files (.pickle), compare them and compute metrics.