Skip to content

AutoLoc: Weakly-supervised Temporal Action Localization in Untrimmed Videos. ECCV'18.

License

Notifications You must be signed in to change notification settings

zhengshou/AutoLoc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

63 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AutoLoc

Citing

If you find AutoLoc or pre-extracted features useful, please consider citing:

@InProceedings{zheng_eccv18_autoloc,
  title={AutoLoc: Weakly-supervised Temporal Action Localization in Untrimmed Videos},
  author={Shou, Zheng and Gao, Hang and Zhang, Lei and Miyazawa, Kazuyuki and Chang, Shih-Fu},
  booktitle = {ECCV},
  year={2018}
}

Installation

This code has been tested with NVIDIA Titan X GPU of 12GB memory, Ubuntu 14.04, CUDA 8.0, cuDNN v5, caffe python interface, python2, python libs specified in requirements.txt (pip install -r requirements.txt). Please use "Issues" to ask questions or report bugs. Thanks.

For caffe, originally I was using MS internal caffe, but in order to publish this code repo, I have tested it with bvlc official caffe. I modified its python interface a little bit regarding the logging part and my customized bvlc caffe has been provided in this repo as well. Note that in order to install it, just need to follow the instructions on the official bvlc caffe website. Please refer any issues of setting up caffe on their venue. Basically the needed procedure is:

  1. cd bvlc_caffe;
  2. prepare your Makefile.config with WITH_PYTHON_LAYER=1 enabled;
  3. make all;
  4. make pycaffe;
  5. export PYTHONPATH=/your_bvlc_caffe_path/python:$PYTHONPATH.

Feature extraction

If you are willing to extract features by yourself, note that for experiments on AN v1.2, please refer to wanglimin/UntrimmedNet#16 to find UntrimmedNet pre-trained model on train set. Models released here http://mmlab.siat.ac.cn/untrimmednet_model/ were trained on train+val set.

Alternatively, we release our extracted features for public downloading.

TH'14: OneDrive or BaiduCloud;

AN v1.2: OneDrive split00, split01 or BaiduCloud split00-split11

Details of the above features:

  1. Put TH14.tar.gz under ./data/TH14/; cd ./data/TH14/; tar -xzvf TH14.tar.gz -C ./; features are contained in the att_unfused/ folder. In att_unfused/:
    • Each video has one individual tsv file storing features.
    • Each row in tsv represents features of one whole video. Each row sequentially contains: video id, class label, feature, class activation, attention score.
    • For feature, class activations, attention score, values are separated by ; and data corresponding to each frame are concatenated sequentially.
    • For feature, the data field has dimension num of frames (T) x dim of feature (2048). T is the total number of frames of the whole video. For the 2048-dim vector, the first half 1024 is corresponding to the temporal CNN and the second half 1024 is spatial.
    • Likewise, class activation has dimension num of frames (T) x 2 x num of classes (K). The first half K is temporal (score after softmax over classes) and the second half is spatial. Attention score has dimension num of frames (T) x 2 (temporal attention, spatial attention).
    • All videos within each subset have stacked into one whole file, i.e. test.all.tsv and val.all.tsv.
  2. For AN:
    • Due to the storage limit, on OneDrive, I have first splited AN.tar.gz into AN.tar.gz.split00 and AN.tar.gz.split01 using split -d -b 19G AN.tar.gz AN.tar.gz.split. Thus after downloading the above splited files, the user needs to first merge them via cat AN.tar.gz.split00 AN.tar.gz.split01 > AN.tar.gz; for Baidu Cloud, more splits (each of 3G) have been used.
    • Put AN.tar.gz under ./data/AN/; cd ./data/AN/; tar -xzvf AN.tar.gz -C ./; features are contained in the att_unfused_onlytrain/ folder
  3. We follow UntrimmedNet to extract feature every 15 frames.

Overview of AutoLoc code structure

  1. data/: contains TH'14 and AN datasets
  2. exp/: scripts for running experiments and our trained model and testing results etc.
  3. lib/: core code of AutoLoc.
  4. tools/: tools for launching AutoLoc and evaluating results. Note that proc_net.py takes a whole set of videos as input and thus usually is used for training. proc_prll_net.py can launch multiple concurrent processes that each of them only takes one single video as input and thus usually is used for testing so that making predictions for each video in parallel to speedup.

Running AutoLoc

Under the root directory of this repo, run python exp/TH14/release/run.py for experiments on TH'14 and python exp/AN/release/run.py for experiments on AN v1.2. These scripts run training first and then run testing. In these scripts, num_process can be used to set the number of processes to be launched during the parallel testing.

Results

THUMOS'14 train on the val set and test on the testing set (mAP):

IoU 0.3 0.4 0.5 0.6 0.7
Sun et al. 8.5 5.2 4.4 - -
Hide-and-Seek 19.5 12.7 6.8 - -
UntrimmedNet 28.2 21.1 13.7 - -
AutoLoc (Ours) 35.8 29.0 21.2 13.4 5.8

AN v1.2 train on the trainig set and test on the val set (mAP):

IoU 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 Avg
UntrimmedNet 7.4 6.1 5.2 4.5 3.9 3.2 2.5 1.8 1.2 0.7 3.6
AutoLoc (Ours) 27.3 24.9 22.5 19.9 17.5 15.1 13.0 10.0 6.8 3.3 16.0

About

AutoLoc: Weakly-supervised Temporal Action Localization in Untrimmed Videos. ECCV'18.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published