The code is synced with the code that is released in our lab's github: https://github.com/Sha-Lab/CMHSE

Cross-Modal and Hierarchical Modeling of Video and Text

The code repository for "Cross-Modal and Hierarchical Modeling of Video and Text" in PyTorch

Prerequisites

The following packages are required to run the scripts:

• PyTorch >= 0.4 and torchvision

• Package tensorboardX and NLTK

• Dataset: please download features and put them into the folder data/anet_precomp and data/didemo_precomp respectively.

• Warning The data is extremly large. It might takes a while for downloading. C3D and ICEP for ActivityNet are ~30G and ~60G, respectively. ICEP for Didemo is ~60G.

Model Evaluation

The learned model on ActivityNet and DiDeMo can be found in this link. You can run train.py with option --resume and --eval_only to evaluate a given model, with options similar to the training scripts as below.

For a model with Inception feature on ActivityNet dataset at "./runs/release/activitynet/ICEP/hse_tau5e-4/run1/checkpoint.pth.tar", it can be evaluated by:

$ python train.py anet_precomp --feat_name icep --img_dim 2048 --resume ./runs/release/activitynet/ICEP/hse_tau5e-4/run1/checkpoint.pth.tar --eval_only

For a model with C3D feature on ActivityNet dataset at "./runs/release/activitynet/C3D/hse_tau5e-4/run1/checkpoint.pth.tar", it can be evaluated by:

$ python train.py anet_precomp --feat_name c3d --img_dim 500 --resume ./runs/release/activitynet/C3D/hse_tau5e-4/run1/checkpoint.pth.tar --eval_only

We presume the input model is a GPU stored model.

Model Training

To reproduce our experiments with HSE, please use train.py and follow the instructions below. We reported the results at the 15th epoch. To train HSE with \tau=5e-4, please with

$ --reconstruct_loss --lowest_reconstruct_loss

For example, to train HSE with \tau=5e-4 on ActivityNet with C3D feature:

$ python train.py anet_precomp --feat_name c3d --img_dim 500 --low_level_loss --reconstruct_loss --lowest_reconstruct_loss --norm

To train HSE with \tau=5e-4 on ActivityNet with Inception feature:

$ python train.py anet_precomp --feat_name icep --img_dim 2048 --low_level_loss --reconstruct_loss --lowest_reconstruct_loss --norm

To train HSE with \tau=5e-4 on Didemo with Inception feature:

$ python train.py didemo_precomp --feat_name icep --img_dim 2048 --low_level_loss --reconstruct_loss --lowest_reconstruct_loss --norm

To train HSE with \tau=0 on ActivityNet with C3D feature:

$ python train.py anet_precomp --feat_name c3d --img_dim 500 --low_level_loss --norm

.bib citation

If this repo helps in your work, please cite the following paper:

@inproceedings{DBLP:conf/eccv/ZhangHS18,
  author    = {Bowen Zhang and
           Hexiang Hu and
           Fei Sha},
  title     = {Cross-Modal and Hierarchical Modeling of Video and Text},
  booktitle = {Computer Vision - {ECCV} 2018 - 15th European Conference, Munich,
           Germany, September 8-14, 2018, Proceedings, Part {XIII}},
  pages     = {385--401},
  year      = {2018}

}

Acknowledgment

We thank following repos providing helpful components/functions in our work.

• VSE++ for the framework
• TSN for the inception-v3 feature

Contacts

Please report bugs and errors to

Bowen Zhang: zbwglory [at] gmail.com
Hexiang Hu: hexiang.frank.hu [at] gmail.com