VoV3D is an efficient and effective 3D backbone network for temporal modeling implemented on top of PySlowFast.


Video classification researches that have recently attracted attention are the fields of temporal modeling and 3D efficient architecture. However, the temporal modeling methods are not efficient or the 3D efficient architecture is less interested in temporal modeling. For bridging the gap between them, we propose an efficient temporal modeling 3D architecture, called VoV3D, that consists of a temporal one-shot aggregation (T-OSA) module and depthwise factorized component, D(2+1)D. The T-OSA is devised to build a feature hierarchy by aggregating temporal features with different temporal receptive fields. Stacking this T-OSA enables the network itself to model short-range as well as long-range temporal relationships across frames without any external modules. Inspired by kernel factorization and channel factorization, we also design a depthwise spatiotemporal factorization module, named, D(2+1)D that decomposes a 3D depthwise convolution into two spatial and temporal depthwise convolutions for making our network more lightweight and efficient. By using the proposed temporal modeling method (T-OSA), and the efficient factorized component (D(2+1)D), we construct two types of VoV3D networks, VoV3D-M and VoV3D-L. Thanks to its efficiency and effectiveness of temporal modeling, VoV3D-L has 6x fewer model parameters and 16x less computation, surpassing a state-of-the-art temporal modeling method on both Something-Something and Kinetics-400. Furthermore, VoV3D shows better temporal modeling ability than a state-of-the-art efficient 3D architecture, X3D having comparable model capacity. We hope that VoV3D can serve as a baseline for efficient video classification.

Main Result

Our results (X3D & VoV3D) are trained in the same environment.

  • V100 8 GPU machine
  • same training protocols (BASE_LR, LR_POLICY, batch size, etc)
  • pytorch 1.6
  • CUDA 10.1

*Please refer to our paper or configs files for the details.
*When you want to reproduce the same results, you just train the model with configs on the 8 GPU machine.
If you change NUM_GPUS or TRAIN.BATCH_SIZE values, you have to adjust BASE_LR.
*IM and K-400 denote ImageNet and Kinetics-400, respectively.


Model Backbone Pretrain #Frame Param. GFLOPs Top-1 Top-5 weight
TSM R-50 K-400 16 24.3M 33x6 48.3 78.1 link
TSM+TPN R-50 IM 8 N/A N/A 50.7 - link
TEA R-50 IM 16 24.4M 70x30 52.3 81.9 -
ip-CSN-152 - - 32 29.7M 74.0x10 49.3 - -
X3D M - 16 3.3M 6.1x6 46.4 75.3 link
VoV3D M - 16 3.3M 5.7x6 48.1 76.9 link
VoV3D M - 32 3.3M 11.5x6 49.8 78.0 link
VoV3D M K-400 32 3.3M 11.5x6 52.6 80.4 link
X3D L - 16 5.6M 9.1x6 47.0 76.4 link
VoV3D L - 16 5.8M 9.3x6 49.5 78.0 link
VoV3D L - 32 5.8M 20.9x6 50.6 78.7 link
VoV3D L K-400 32 5.8M 20.9x6 54.9 82.3 link


Model Backbone Pretrain #Frame Param. GFLOPs Top-1 Top-5 weight
TSM R-50 K-400 16 24.3M 33x6 63.0 88.1 link
TSM+TPN R-50 IM 8 N/A N/A 64.7 - link
TEA R-50 IM 16 24.4M 70x30 65.1 89.9 -
SlowFast 16x8 R-50 K-400 64 34.0M 131.4x6 63.9 88.2 link
X3D M - 16 3.3M 6.1x6 63.0 87.9 link
VoV3D M - 16 3.3M 5.7x6 63.2 88.2 link
VoV3D M - 32 3.3M 11.5x6 64.2 88.8 link
VoV3D M K-400 32 3.3M 11.5x6 65.2 89.4 link
X3D L - 16 5.6M 9.1x6 62.7 87.7 link
VoV3D L - 16 5.8M 9.3x6 64.1 88.6 link
VoV3D L - 32 5.8M 20.9x6 65.8 89.5 link
VoV3D L K-400 32 5.8M 20.9x6 67.3 90.5 link


Model Backbone Pretrain #Frame Param. GFLOPs Top-1 Top-5 weight
X3D (PySlowFast, 300e) M - 16 3.8M 6.2x30 76.0 92.3 link
X3D (our, 256e) M - 16 3.8M 6.2x30 75.0 92.1 link
VoV3D M - 16 3.8M 4.4x30 73.9 91.6 link
X3D (PySlowfast) L - 16 6.1M 24.8x30 77.5 92.9 link
VoV3D L - 16 6.2M 9.3x30 76.3 92.9 link

*We note that since X3D-M (PySlowFast) was trained for 300 epochs, we re-train the X3D-M (our, 256e) with the same 256 epochs with VoV3D-M.

Installation & Data Preparation

Please refer to INSTALL.md for installation and DATA.md for data preparation.
Important : We used depthwise 3D Conv pytorch patch for accelearating GPU runtime.

Training & Evaluation

We provide brief examples for getting started. If you want to know more details, please refer to instruction of PySlowFast.


from scratch

  • VoV3D-L on Kinetics-400
python tools/run_net.py \
  --cfg configs/Kinetics/vov3d/vov3d_L.yaml \
  DATA.PATH_TO_DATA_DIR path/to/your/kinetics \
  NUM_GPUS 8 \

You can also designate each argument in the config file. If you want to train with our default setting (e.g., 8GPUs, 64 batch size, etc), you just use this command. (Set DATA.PATH_TO_DATA_DIR with your real data path)

python tools/run_net.py --cfg configs/Kinetics/vov3d/vov3d_L.yaml
  • VoV3D-L on Something-Something-V1
python tools/run_net.py \
  --cfg configs/SSv1/vov3d/vov3d_L_F16.yaml \
  DATA.PATH_TO_DATA_DIR path/to/your/ssv1 \ 
  DATA.PATH_PREFIX path/to/your/ssv1

Finetuning by using Kinetics-400 pretrained weight.

First, you have to download the weights pretrained on Kinetics-400.

One thing you should keep in mind is that TRAIN.CHECKPOINT_FILE_PATH is the downloaded weight.

For Something-Something-V2,

cd VoV3D
mkdir -p output/pretrained
wget https://dl.dropbox.com/s/lzmq8d4dqyj8fj6/vov3d_L_k400.pth

python tools/run_net.py \
  --cfg configs/SSv2/vov3d/finetune/vov3d_L_F16.yaml \
  TRAIN.CHECKPOINT_FILE_PATH path/to/the/pretrained/vov3d_L_k400.pth \
  DATA.PATH_TO_DATA_DIR path/to/your/ssv2 \
  DATA.PATH_PREFIX path/to/your/ssv2


When testing, you have to set TRAIN.ENABLE to False and TEST.CHECKPOINT_FILE_PATH to path/to/your/checkpoint.

python tools/run_net.py \
  --cfg configs/Kinetics/vov3d/vov3d_L.yaml \
  TEST.CHECKPOINT_FILE_PATH path_to_your_checkpoint

If you want to test with single clip and single-crop, set TEST.NUM_ENSEMBLE_VIEWS and TEST.NUM_SPATIAL_CROPS to 1, respectively.

python tools/run_net.py \
  --cfg configs/Kinetics/vov3d/vov3d_L.yaml \
  TEST.CHECKPOINT_FILE_PATH path_to_your_checkpoint \

For Kinetics-400, 30-views : TEST.NUM_ENSEMBLE_VIEWS 10 & TEST.NUM_SPATIAL_CROPS 3
For Something-Something, 6-views : TEST.NUM_ENSEMBLE_VIEWS 2 & TEST.NUM_SPATIAL_CROPS 3


The code and the models in this repo are released under the CC-BY-NC4.0 LICENSE. See the LICENSE file.

Citing VoV3D

  title={Diverse Temporal Aggregation and Depthwise Spatiotemporal Factorization for Efficient Video Classification},
  author={Lee, Youngwan and Kim, Hyung-Il and Yun, Kimin and Moon, Jinyoung},
  journal={arXiv preprint arXiv:2012.00317},

  title = {An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection},
  author = {Lee, Youngwan and Hwang, Joong-won and Lee, Sangrok and Bae, Yuseok and Park, Jongyoul},
  booktitle = {CVPR Workshop},
  year = {2019}

  title={CenterMask: Real-Time Anchor-Free Instance Segmentation},
  author={Lee, Youngwan and Park, Jongyoul},


We appreciate developers of PySlowFast for such wonderful framework.
This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. B0101-15-0266, Development of High Performance Visual BigData Discovery Platform for Large-Scale Realtime Data Analysis and No. 2020-0-00004, Development of Previsional Intelligence based on Long-term Visual Memory Network).