Pytorch implementation of paper "Marginalized Average Attentional Network for Weakly-Supervised Learning" (ICLR 2019).

[PDF]

Code is tested with Pytorch 1.0 + Python 3.6.

A simple test for the marginalized average aggregation (MAA) layer is provided with:

python MAA.py

The aggregator is initially designed to aggregate video features, i.e. input is a 3D tensor : Batch Size X T X N

But it can easily be adapted to process image feature map, you only need to reshape 4D tensors into 3D tensors.

Unlike classical attentional aggregator, MAA takes the expectation of the average aggregated subsetfeatures over all the possible subsets to achieve the final aggregation.

The pipeline for video action localization under weakly supervised setting (only video level action label available) is illustrated below :

Some visual results :

If our project is helpful for your research, please consider citing :

@inproceedings{yuan2018marginalized,
          title={Marginalized Average Attentional Network for Weakly-Supervised Learning},
          author={Yuan, Yuan and Lyu, Yueming and Shen, Xi and Tsang, Ivor W and Yeung, Dit-Yan},
          booktitle={International Conference on Learning Representations (ICLR)},
          year={2019}
        }