levilian/actnet-inact

This repository is intended to host tools and demos for ActivityNet

Introduction

This repository was intended to host the work I did during ActivityNet-2016 challenge. We received 2nd prize in detection task of the challenge. Also, this secured 1st position on average mAP. Also, it was positioned at the 10th place in the classification task.

This work described in a technical report available at arxiv.

Please refer to Evaluation page of the [challenge] (http://activity-net.org/challenges/2016/evaluation.html) to check the standing.

Repository Structure

This repository is built upon the fork of activitynet/ActivityNet repo.

Crawler: is modified for personal use, you can use it from original [repo] as well.

python-script: contains the script to convert .mp4 video int .jpg frames produce the level labels. Originally temporal labelling is done in from start time and end time in 'seconds'. We resave these annotations in frame level labels.

There are two scripts to extract frames. Matlab script implement the functionality of reading the video and converting it into frames python script has the same ability and much faster in doing that

Then, there is a python script to re-save the annotation according to frame level annotations and action class ID; action class IDs are from 1-200 continuous number, unlike nodeId that is provided. This conversion is reversible as well.

processing: contains python script to train SVM on the different kind of features and save frame level score vector for each frame of validation and testing videos.

Further, there are two main scripts to produce classification and detection results. Detection script contains the implementation of dynamic programming based solver to solve labelling problem initially proposed in [1].

Citing

If you find this work useful in your research, please consider citing:

@article{singh2016untrimmed,
   title={Untrimmed Video Classification for Activity Detection: submission to ActivityNet Challenge},
   author={Singh, Gurkirt and Cuzzolin, Fabio},
   journal={arXiv preprint arXiv:1607.01979},
   year={2016}
 }

This code is released without any guarantee.

Reference

[1] Georgios, Evangelidis, Gurkirt Singh and Radu Horaud. "Continuous gesture recognition from articulated poses." Workshop at the European Conference on Computer Vision. Springer International Publishing, 2014.