Awesome

BaSSL

This is an official PyTorch Implementation of Boundary-aware Self-supervised Learning for Video Scene Segmentation (BaSSL) [arxiv] [demo in modelscope]

The method is a self-supervised learning algorithm that learns a model to capture contextual transition across boundaries during the pre-training stage. To be specific, the method leverages pseudo-boundaries and proposes three novel boundary-aware pretext tasks effective in maximizing intra-scene similarity and minimizing inter-scene similarity, thus leading to higher performance in video scene segmentation task.

1. Environmental Setup

We have tested the implementation on the following environment:

Python 3.7.7 / PyTorch 1.7.1 / torchvision 0.8.2 / CUDA 11.0 / Ubuntu 18.04

Also, the code is based on pytorch-lightning (==1.3.8) and all necessary dependencies can be installed by running following command.

$ pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
$ pip install -r requirements.txt

# (optional) following installation of pillow-simd sometimes brings faster data loading.
$ pip uninstall pillow && CC="cc -mavx2" pip install -U --force-reinstall pillow-simd

2. Prepare Data

We provide data download script for raw key-frames of MovieNet-SSeg dataset, and our re-formatted annotation files applicable for BaSSL. FYI, our script will automatically download and decompress data---1) key-frames (160G), 2) annotations (200M)---into <path-to-root>/bassl/data/movienet.

# download movienet data
$ cd <path-to-root>
$ bash script/download_movienet_data.sh

In addition, download annotation files from MovieNet-SSeg google drive and put the folder scene318 into <path-to-root>/bassl/data/movienet. Then, the data folder structure will be as follows:

# <path-to-root>/bassl/data
movienet
│─ 240P_frames
│    │─ tt0120885                 # movie id (or video id)
│    │    │─ shot_0000_img_0.jpg
│    │    │─ shot_0000_img_1.jpg
│    │    │─ shot_0000_img_2.jpg  # for each shot, three key-frames are given.
|    |    :
│    :    │─ shot_1256_img_2.jpg
│    |    
│    │─ tt1093906
│         │─ shot_0000_img_0.jpg
│         │─ shot_0000_img_1.jpg
│         │─ shot_0000_img_2.jpg
|         :
│         │─ shot_1270_img_2.jpg
│
│─anno
     │─ anno.pretrain.ndjson
     │─ anno.trainvaltest.ndjson
     │─ anno.train.ndjson
     │─ anno.val.ndjson
     │─ anno.test.ndjson
     │─ vid2idx.json
│─scene318
     │─ label318
     │─ meta
     │─ shot_movie318

3. Train (Pre-training and Fine-tuning)

We use Hydra to provide flexible training configurations. Below examples explain how to modify each training parameter for your use cases.
We assume that you are in <path-to-root> (i.e., root of this repository).

3.1. Pre-training

(1) Pre-training BaSSL
Our pre-training is based on distributed environment (multi-GPUs training) using ddp environment supported by pytorch-lightning.
The default setting requires 8-GPUs (of V100) with a batch of 256. However, you can set the parameter config.DISTRIBUTED.NUM_PROC_PER_NODE to the number of gpus you can use or change config.TRAIN.BATCH_SIZE.effective_batch_size. You can run a single command cd bassl; bash ../scripts/run_pretrain_bassl.sh or following full command:

cd <path-to-root>/bassl
EXPR_NAME=bassl
WORK_DIR=$(pwd)
PYTHONPATH=${WORK_DIR} python3 ${WORK_DIR}/pretrain/main.py \
    config.EXPR_NAME=${EXPR_NAME} \
    config.DISTRIBUTED.NUM_NODES=1 \
    config.DISTRIBUTED.NUM_PROC_PER_NODE=8 \
    config.TRAIN.BATCH_SIZE.effective_batch_size=256

Note that the checkpoints are automatically saved in bassl/pretrain/ckpt/<EXPR_NAME> and log files (e.g., tensorboard) are saved in `bassl/pretrain/logs/<EXPR_NAME>.

(2) Running with various loss combinations
Each objective can be turned on and off independently.

cd <path-to-root>/bassl
EXPR_NAME=bassl_all_pretext_tasks
WORK_DIR=$(pwd)
PYTHONPATH=${WORK_DIR} python3 ${WORK_DIR}/pretrain/main.py \
    config.EXPR_NAME=${EXPR_NAME} \
    config.LOSS.shot_scene_matching.enabled=true \
    config.LOSS.contextual_group_matching.enabled=true \
    config.LOSS.pseudo_boundary_prediction.enabled=true \
    config.LOSS.masked_shot_modeling.enabled=true

(3) Pre-training shot-level pre-training baselines
Shot-level pre-training methods can be trained by setting config.LOSS.sampling_method.name as one of followings:

instance (Simclr_instance), temporal (Simclr_temporal), shotcol (Simclr_NN).
And, you can choose two more options: bassl (BaSSL), and bassl+shotcol (BaSSL+ShotCoL).
Below example is for Simclr_NN, i.e., ShotCoL. Choose your favorite option ;)

cd <path-to-root>/bassl
EXPR_NAME=Simclr_NN
WORK_DIR=$(pwd)
PYTHONPATH=${WORK_DIR} python3 ${WORK_DIR}/pretrain/main.py \
    config.EXPR_NAME=${EXPR_NAME} \
    config.LOSS.sampleing_method.name=shotcol \

3.2. Fine-tuning

(1) Simple running a single command to fine-tune pre-trained models
Firstly, download the checkpoints provided in Model Zoo section and move them into bassl/pretrain/ckpt.

cd <path-to-root>/bassl

# for fine-tuning BaSSL (10 epoch)
bash ../scripts/finetune_bassl.sh

# for fine-tuning Simclr_NN (i.e., ShotCoL)
bash ../scripts/finetune_shot-level_baseline.sh

The full process (i.e., extraction of shot-level representation followed by fine-tuning) is described in below.

(2) Extracting shot-level features from shot key-frames
For computational efficiency, we pre-extract shot-level representation and then fine-tune pre-trained models.
Set LOAD_FROM to EXPR_NAME used in the pre-training stage and change config.DISTRIBUTED.NUM_PROC_PER_NODE as the number of GPUs you can use. Then, the extracted shot-level features are saved in <path-to-root>/bassl/data/movienet/features/<LOAD_FROM>.

cd <path-to-root>/bassl
LOAD_FROM=bassl
WORK_DIR=$(pwd)
PYTHONPATH=${WORK_DIR} python3 ${WORK_DIR}/pretrain/extract_shot_repr.py \
	config.DISTRIBUTED.NUM_NODES=1 \
	config.DISTRIBUTED.NUM_PROC_PER_NODE=1 \
	+config.LOAD_FROM=${LOAD_FROM}

(3) Fine-tuning and evaluation

cd <path-to-root>/bassl
WORK_DIR=$(pwd)

# Pre-training methods: bassl and bassl+shotcol
# which learn CRN network during the pre-training stage
LOAD_FROM=bassl
EXPR_NAME=transfer_finetune_${LOAD_FROM}
PYTHONPATH=${WORK_DIR} python3 ${WORK_DIR}/finetune/main.py \
	config.TRAIN.BATCH_SIZE.effective_batch_size=1024 \
	config.EXPR_NAME=${EXPR_NAME} \
	config.DISTRIBUTED.NUM_NODES=1 \
	config.DISTRIBUTED.NUM_PROC_PER_NODE=1 \
	config.TRAIN.OPTIMIZER.lr.base_lr=0.0000025 \
	+config.PRETRAINED_LOAD_FROM=${LOAD_FROM}

# Pre-training methods: instance, temporal, shotcol
# which DO NOT learn CRN network during the pre-training stage
# thus, we use different base learning rate (determined after hyperparameter search)
LOAD_FROM=shotcol_pretrain
EXPR_NAME=finetune_scratch_${LOAD_FROM}
PYTHONPATH=${WORK_DIR} python3 ${WORK_DIR}/finetune/main.py \
	config.TRAIN.BATCH_SIZE.effective_batch_size=1024 \
	config.EXPR_NAME=${EXPR_NAME} \
	config.DISTRIBUTED.NUM_NODES=1 \
	config.DISTRIBUTED.NUM_PROC_PER_NODE=1 \
	config.TRAIN.OPTIMIZER.lr.base_lr=0.000025 \
	+config.PRETRAINED_LOAD_FROM=${LOAD_FROM}

4. Model Zoo

We provide pre-trained checkpoints trained in a self-supervised manner.
After fine-tuning with the checkpoints, the models will give scroes that are almost similar to ones shown below.

Method	AP	Checkpoint (pre-trained)
SimCLR (instance)	51.51	download
SimCLR (temporal)	50.05	download
SimCLR (NN)	51.17	download
BaSSL (10 epoch)	56.26	download
BaSSL (40 epoch)	57.40	download

5. Citation

If you find this code helpful for your research, please cite our paper.

@article{mun2022boundary,
  title={Boundary-aware Self-supervised Learning for Video Scene Segmentation},
  author={Mun, Jonghwan and Shin, Minchul and Han, Gunsu and
          Lee, Sangho and Ha, Sungsu and Lee, Joonseok and Kim, Eun-sol},
  journal={arXiv preprint arXiv:2201.05277},
  year={2022}
}

6. Contact for Issues

Jonghwan Mun, jason.mun@kakaobrain.com
Minchul Shin, craig.starr@kakaobrain.com