• 设为首页
  • 点击收藏
  • 手机版
    手机扫一扫访问
    迪恩网络手机版
  • 关注官方公众号
    微信扫一扫关注
    迪恩网络公众号

HYPJUDY/Decouple-SSAD: Decoupling Localization and Classification in Single Shot ...

原作者: [db:作者] 来自: 网络 收藏 邀请

开源软件名称(OpenSource Name):

HYPJUDY/Decouple-SSAD

开源软件地址(OpenSource Url):

https://github.com/HYPJUDY/Decouple-SSAD

开源编程语言(OpenSource Language):

Python 86.9%

开源软件介绍(OpenSource Introduction):

Decouple-SSAD

Code, models and data for Decouple-SSAD.

Decoupling Localization and Classification in Single Shot Temporal Action Detection,
Yupan Huang, Qi Dai, Yutong Lu, in ICME 2019.

Download preprocessed data, final models and the result files for evaluation by visiting Onedrive.

An improved version of SSAD (Single Shot Temporal Action Detection) is also provided in this repository.

Introduction

Video temporal action detection aims to temporally localize and recognize the action in untrimmed videos. Existing one-stage approaches mostly focus on unifying two subtasks, i.e., localization of action proposals and classification of each proposal through a fully shared backbone. However, such design of encapsulating all components of two subtasks in one single network might restrict the training by ignoring the specialized characteristic of each subtask. In this paper, we propose a novel Decoupled Single Shot temporal Action Detection (Decouple-SSAD) method to mitigate such problem by decoupling the localization and classification in a one-stage scheme. Particularly, two separate branches are designed in parallel to enable each component to own representations privately for accurate localization or classification. Each branch produces a set of action anchor layers by applying deconvolution to the feature maps of the main stream. High-level semantic information from deeper layers is thus incorporated to enhance the feature representations. We conduct extensive experiments on THUMOS14 and demonstrate superior performance over state-of-the-art methods.

Environment

  • TensorFlow
  • One or more GPU with 12G memory
  • Some packages, e.g. pandas, numpy

Have been tested on Ubuntu 16.04

  • Python2.7, cuda-8.0, tensorflow v1.4.1
  • Python3.5/3.6, cuda-9.0, tensorflow v1.12.0

For example, I use the following commands to set up environment:

export PATH=/usr/local/cuda-9.0/bin:$PATH  # select cuda version
export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64/

conda create -n Decouple-SSAD pip python=3.6  # select python version
source activate Decouple-SSAD
pip install --upgrade packageURL  # select tensorflow version
# e.g. tensorflow_gpu-1.12.0-cp36-cp36m-linux_x86_64.whl
# download from https://www.tensorflow.org/install/pip#package-location
# or
# pip install tensorflow-gpu --user 
# this command will install the latest version, which may not match with your cuda version

git clone https://github.com/HYPJUDY/Decouple-SSAD.git
cd Decouple-SSAD

# ... follow "Run code" section ...

conda deactivate

Run code

  1. Prepare THUMOS14 data.
  2. (Optional) Download final models for direct evaluation.
  3. (Optional) Specify your path in config.py (e.g. feature_path, get_models_dir(), get_predict_result_path()).
  4. (Optional) If you want to experiment with a small dataset to save time, set small_num_data_for_test = True in load_data.py.
  5. Modify run.sh to specify gpu(s) device and the other parameters according to your need.
    bash ./run.sh # for training, evaluation or two-stream results fusion
    The default run.sh do the following things:
    1. run the two process in parallel:
      • Train and then test Decouple-SSAD pretrained on UCF101 of temporal mode in GPU 0
      • Train and then test Decouple-SSAD pretrained on UCF101 of spatial mode in GPU 1
    2. Then the temporal and spatial streams are fused and results will be saved at results folder (e.g., results/result_fuse_UCF101_decouple_ssad.txt). Training logs are saved at logs folder. Models are save at models folder.
  6. I evaluate the results by official Matlab code
    cd EvalKit/THUMOS14_evalkit_20150930
    [pr_all,ap_all,map]=TH14evalDet('Decouple-SSAD/results/result_fuse_UCF101_decouple_ssad.txt','annotation','test',0.5);
    

Compare with SSAD

Our implementation of Decouple-SSAD is based on SSAD (Single Shot Temporal Action Detection). We thank Tianwei Lin for his generously sharing code. We re-train the origin SSAD as our baseline model (i.e., akin to the single main stream of Decouple-SSAD), which follows the overall structure and adopt most of the parameters of SSAD but with the following improvements:

  1. We use different video features as model input. SSAD use Snippet-level Action Score (SAS) feature of 303 dimension by concatenating the output scores of action classifiers in spatial, temporal and C3D network. While we use the output of feature layer in defined network (e.g. ResNet) as feature. Spatial and temporal features of 1024/2048 dimensions are input to Decouple-SSAD separately and fused lately.
  2. The main hyper-parameters we changed are:
    • the number of convolutional filter in all base layers: 256 -> 512
    • the number of convolutional filter in multi-scale anchor layers: 512 -> 1024
    • scale ratios: {1,1.5,2} for fA1 and {0.5,0.75,1,1.5,2} for fA2 and fA3 -> {0.5,0.75,1,1.5,2} for fA1, fA2 and fA3
  3. In post-processing, we assign cliffDiving class as diving class too

Besides to these, some codes of SSAD have been re-structured and modified, which may also have influence on the performance. We conduct ablation study to show that our baseline model (the main stream pretrained on UCF101) improve mAP@0.5 performance of SSAD from 24.6% to 31.2% (6.6% relative gains). Moreover, Decouple-SSAD could further enhance the performance significantly, providing 4.6% enhancement against the baseline.

Performance

Note that the performance may fluctuate 0.5~2% at each training even with the same environment and code. Occasionally, the mAP of single temporal stream is higher than the fused result (e.g. Decouple-SSAD(KnetV3) in my last training). Here's the detailed performance report. Each of the following experiments are run with 2 GeForce RTX 2080 Ti GPUs.

Method pre-train fuse mAP@0.5 temporal mAP@0.5 spatial mAP@0.5 train, test and fuse time single-stream training time
Decouple-SSAD KnetV3 43.7% 44.2% 30.7% 86min 73min
Decouple-SSAD UCF101 35.8% 33.1% 22.1% 77min 65min
main-stream-SSAD UCF101 31.2% 30.0% 18.3% 53min 42min

For those who are interested in the detection AP over different action categories with overlap threshold 0.5 in THUMOS14, please refer to fig/plot_class_ap.py for precised number and a visualization comparision is as follow:

Citation

If you like this paper or code, please cite us:

@inproceedings{huang2019decoupling,
  author    = {Yupan Huang and Qi Dai and Yutong Lu},
  title     = {Decoupling Localization and Classification in Single Shot Temporal Action Detection},
  booktitle = {International Conference on Multimedia and Expo (ICME)},
  year      = {2019}
}

Contact

Feel free to open issues or email to me for issues or help using Decouple-SSAD. Any feedback is welcome!




鲜花

握手

雷人

路过

鸡蛋
该文章已有0人参与评论

请发表评论

全部评论

专题导读
热门推荐
阅读排行榜

扫描微信二维码

查看手机版网站

随时了解更新最新资讯

139-2527-9053

在线客服(服务时间 9:00~18:00)

在线QQ客服
地址:深圳市南山区西丽大学城创智工业园
电邮:jeky_zhao#qq.com
移动电话:139-2527-9053

Powered by 互联科技 X3.4© 2001-2213 极客世界.|Sitemap