By Jingjia Huang, Nannan Li, Tao Zhang and Ge Li

The paper can be found here.

Self-Adaptive Proposal (SAP) is a DQN based model for temporal action localization in untrimmed long videos. The temporal action detection process for SAP is naturally one of observation and refinement: observe the current window and refine the span of attended window to cover true action regions. SAP can learn to find actions through continuously adjusting the temporal bounds in a self-adaptive way. Experiment results on THUMOS’14 validate the effectiveness of SAP, which can achieve competitive performance with current action detection algorithms via much fewer proposals.

fig.1 SAP architecture

fig.2 Illustration of DQN actions.Each yellow window with dashed lines represents the next window after taking the corresponding action.

fig.3 Example of how SAP works

This code has been tested on Ubuntu 16.04 with NVIDIA Tesla K80. The CUDA version is 8.0.61

SAP is released under the MIT License.

If you find SAP useful, please consider citing:

@paper{AAAI1816109,
	author = {Jingjia Huang and Nannan Li and Tao Zhang and Ge Li and Tiejun Huang and Wen Gao},
	title = {SAP: Self-Adaptive Proposal Model for Temporal Action Detection Based on Reinforcement Learning},
	conference = {AAAI Conference on Artificial Intelligence},
	year = {2018},
	keywords = {Computer vision; Action detection; Reinforcement learning},
	url = {https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/16109}
}

If you like this project, give us a ⭐ in the github banner 😉.

0. Ensure that you have gcc, torch7, CUDA and CUDNN.
1. Clone our repo, git https://github.com/hjjpku/Action_Detection_DQN
2. Download the pre-trained c3d v1.0 torch model from
   • https://github.com/wandering007/c3d_torch    - Move the c3d model to our project folder and named it c3d.t7
3. Download pre-trained fc models to our project folder from BaiduYun

• [Pre-trained DQN models]BaiduYun. We provide two version of DQN for the first 4 category of actions. One for the model w/ temporal pooling and another for the model w/o temporal pooling.
• Pre-computed action proposals. Take a look at our results if you are interested in comparisons or building cool algorithms on top of our outputs.

• If you want to train your own models on Thumos'14

  0. Preproceesing for Thumos dataset Down-sample the videos for computing efficiency and save the videos as images. Considering the length of the groundtruth for different action categories varies a lot and the C3D needs an input longer than 16 frames, we down-sample the video of different categories to {1/3;1/3;1;1/2;1;1/3;1/3;1/3;1/3;1/2;1;1/2;1;1/3;1/3;1/3;1/3;1/3;1/3;1/3;1/2} of the original frame rate, respectively.

  1. Construct the dataset directory structure as follow:

   .
   ├── dataset          # put your datasets on the project folder and named it with dataset name
   │   ├── class        # action class name (CleanAndJerk, BaseballPitch, HighJump...)
   │        ├──    
   │        │    ├── clip_idx             # index of videos
   │        │    |   ├── frame_idx.jpg       # video images(from 1 to total sampled image number)
   │        │    |   ...
   │        │    |   ├── frame_idx.jpg
   │        │    |   ├── FrameNum.txt        # total sampled image number
   │        │    |   ├── gt.txt              # groundtruth interval
   │        │    |   ├── NumOfGt.txt         # number of groundtruth in the clip
   │        │    |   └── 
   │        │    ...
   │        │    ├── clip_idx 
   │        │    ├── ClipNum.txt          # number of clips in this category
   │        │    └── 
   │        └── 
   │   ├── class
   |   ...
   |   ├── class
   |   ├── trainlist.t7
   |   ├── validatelist.t7
   └──
  

  2. Perepare the metadata and annotations for training and testing
  • trainlist.t7 and Thumos_validatelis_new.t7 are the data which indicates the index of videos used for training or validation. It should be a table as follow:

  {
    1 --class    
    {
       1 --clip_idx
       ...
    }
    ...
    21 --class
    {
      1 --clip_idx 
      ...
    }
  }
  

  • ClipNum.txt : a single number indicates the number of clips in this category

  • FrameNum.txt : a single number indicates the total sampled image number from the clip

  • NumOfGt.txt : a single number indicates the number of groundtruth segments in the clip

  • gt.txt records the index of first frame and last frame of the groundtruth.It should be arranged as follow:

  <begin_1> <end_1>
  ...
  <begin_n> <end_n>
  

  😉 If you just want to try out our work and don't want to be bothered by this tedious work, you can download the dataset we have already processed here .

  3. Run the code. For example: th ./Hjj_Training_SAP.lua -data_path Thumos -batch_size 200 -replay_buffer 2000 -lr 1e-3 -class 1 to train the SAP model for first action category in dadaset Thumos. If you want to try more simple versions we used for ablation study(SAP w/o temporal pooling; SAP w/o temparal pooling w/o regression), run ./Hjj_Training_DQNRGN.lua or ./Hjj_Trianing_Dqn.lua instead. For more details about the script arguments, please consult Hjj_Read_Input_Cmd.lua