This is an official implementation of AutoTransition: Learning to Recommend Video Transition Effects.
We release the videos with annotated transitions extracted from the video editing template on online video editing platforms. Due to the privacy policy, we only release the link to the videos. The dataset can be downloaded from here: Download Link
Use the following command to download the source video in the dataset:
python3 tools/download_videos.py annotation.json ./template_download
The videos will be downloaded to ./template_download
.
To speed up the training, we convert videos to JPEG image and extract audio features before training. Run the following commands to finish these steps:
python3 preprocess/convert_video_folder.py ./path/to/template_root
python3 preprocess/extract_audio_features.py ./path/to/template_root path/to/annotation.json --model_path /path/to/audio/model.pth --cuda
The pretrained Harmonic CNN model could be downloaded from this link.
To train transition embeddings:
python3 tools/run_net.py --cfg configs/base/train_transition_embedding.yaml \
DATASET.TRANSITION_CLASSIFICATION.JSON_ANNOTATION /path/to/annotation.json \
DATASET.TRANSITION_CLASSIFICATION.TEMPLATE_ROOT /path/to/template_root
The transition embeddings can be found in ./log
directory after training.
To train transition recommendation:
python3 tools/run_net.py --cfg configs/base/train_transition_recommendation.yaml \
MODEL.TRANSITION_TRANSFORMER.EMBEDDING.PRETRAINED_EMBEDDING /path/to/pretrained/transition/embedding.pth \
DATASET.TRANSITION_DATASET.JSON_ANNOTATION /path/to/annotation.json \
DATASET.TRANSITION_DATASET.TEMPLATE_ROOT /path/to/template_root
tensorboard --logdir=./log