Video to Language Challenge (MSR-VTT Challenge 2016)

Team Fudan-ILC (Fudan University & Intel Labs China) solution for the MSR-VTT Challenge (http://ms-multimedia-challenge.com/2016/challenge).

This repository contains the code for the original implementation in the challenge and our updated version.

The original version was ranked 4th in human evaluation and ranked 5th in the automatic evaluation metrics based Leaderboard (Note that we only used a single model (no ensemble) and did not use any category or audio information provided by the dataset in this version.)

The updated version is part of the DenseVidCap project that combines the audio features, category information and other techniques, which achieves higher scores.

This code is based on NeuralTalk2 (https://github.com/karpathy/neuraltalk2).

Table of Contents

1. Usage
2. Train challenge models
3. Test challenge models
4. Challenge results
5. Train updated models
6. Updated results
7. Two improvements for NeuralTalk2
8. Contact

Usage

1. Clone our repository:

git clone https://github.com/szq0214/MSR-VTT-Challenge.git

Requirements (modified from NeuralTalk2)

This code is written in Lua and requires Torch. If you're on Ubuntu, installing Torch in your home directory may look something like:

$ curl -s https://raw.githubusercontent.com/torch/ezinstall/master/install-deps | bash
$ git clone https://github.com/torch/distro.git ~/torch --recursive
$ cd ~/torch; 
$ ./install.sh      # and enter "yes" at the end to modify your bashrc
$ source ~/.bashrc

See the Torch installation documentation for more details. After Torch is installed we need to get a few more packages using LuaRocks (which already came with the Torch install). In particular:

$ luarocks install nn
$ luarocks install nngraph 
$ luarocks install image 

We're also going to need the cjson library so that we can load/save json files. Follow their download link and then look under their section 2.4 for easy luarocks install.

If you'd like to run on an NVIDIA GPU using CUDA (which you really, really want to if you're training a model, since we're using a VGGNet), you'll of course need a GPU, and you will have to install the CUDA Toolkit. Then get the cutorch and cunn packages:

$ luarocks install cutorch
$ luarocks install cunn

If you'd like to use the cudnn backend (the pretrained checkpoint does), you also have to install cudnn. First follow the link to NVIDIA website, register with them and download the cudnn library. Then make sure you adjust your LD_LIBRARY_PATH to point to the lib64 folder that contains the library (e.g. libcudnn.so.7.0.64). Then git clone the cudnn.torch repo, cd inside and do luarocks make cudnn-scm-1.rockspec to build the Torch bindings.

Finally, you will also need to install torch-hdf5, and h5py, since we will be using hdf5 files to store the preprocessed data.

Other preparations

2. Download training and validation features (including the updated version) and put them into root directory of this project. You can also download our pre-trained model here which is used in the challenge.
3. Download the coco-caption code into coco-caption directory.
4. Download the MSR-VTT annotations into coco-caption/annotations directory.

Train challenge models

1. Set training data, json file and options in mean_pool_MSR_VTT_train.lua file. Use -input_h5 data_augmentation.h5, -input_json data_augmentation.json, -rnn_size 1024, -input_encoding_size 1024, -update_iter 4, -learning_rate 4e-4, -beam_size 1, -train_split train, -eval_split val, -val_images_use 497, -save_checkpoint_every 250, -language_eval 1 options.
2. Set feature dimension -input_feature_dim 27648 .

The feature we used includs: 1) ResNet (2048); 2) VGG19 (4096); 3) Places-VGG16 (4096); 4) Places-GoogLeNet (1024); 5) EventNet (4096); 6) C3D_2 (4096); 7) C3D_8 (4096); 8) C3D_16 (4096). The total feature dimension is 27648.

Data augmentation (segment video clip):(1) mean-pooling over all frames from each clip (100%); (2) over first 25% frames; (3) over first 50% frames(4) over first 75% frames; (5) over last 50% frames.

3. Train a model using

th mean_pool_MSR_VTT_train.lua -gpuid 0 | tee MSR_VTT_challenge.log

Test challenge models

1. Set test data, json file and options in eval_MSR_VTT.lua file. Use -input_feature_dim 27648 , -model model_id_MSR_VTT_challenge.t7, -num_images 2990, -language_eval 1, -input_h5 data_augmentation_test.h5, -input_json data_augmentation_test.json, -beam_size 2, -split val options.

2. Test the model using

th eval_MSR_VTT.lua -gpuid 0

Challenge results

The validation score curve during training (Score = BLEU@4 + METEOR + CIDEr + ROUGE-L):

The tables below show the results of Fudan-ILC on MSR-VTT challenge.

M1 performance:

Team	BLEU@4	METEOR	CIDEr	ROUGE-L
Fudan-ILC (validation set)	39.0	27.7	44.0	60.1
Fudan-ILC (test set)	38.7	26.8	41.9	59.5

M2 performance:

Team	C1	C2	C3
Fudan-ILC (test set)	3.185	2.999	2.979

Train updated models

1. Train language models with visual features (category-wise manner)

• Train language models for category_X (replace X below with 0,1,...,19 to train 20 category-wise models)

  1. Set the training data, json file and options. Use -input_feature_dim 11264, -input_h5 data_lexical.h5, -input_json data_lexical.json, -rnn_size 512 -update_iter 1, -learning_rate 2e-4, -beam_size 1, -train_split train_X, -eval_split val_X, -val_images_use XXX, -save_checkpoint_every 100, -checkpoint_path lexical, -language_eval 1, -id _lexical_X options.

  If you want to evaluate the whole validation set, please make -val_images_use larger than the number of examples in each category. For convenience, you can set it with a large number like 1000 for all categories.

  In the challenge model, we use two linear layers to embed the input features, while for efficiency, we apply single layer in the updated model. The embedding parameters are learned jointly with the language model. You can modify line 24~31 in misc/net_utils.lua to

  cnn_part:add(nn.Linear(opt.input_feature_dim, encoding_size, true))

  cnn_part:add(backend.ReLU(true))

  cnn_part:add(nn.Dropout(p2))

  2. Train a model using

  th mean_pool_MSR_VTT_train.lua -gpuid 0 | tee lexical/lexical_X.log

• Calculate the final results: Since each category has different numbers of videos, we can not simply average all best performance scores of all categories. We need to collect all generated best-sentences into a single file from lexical_X.log files (You can search 't7' to find out the best sentences (with highest scores) for collecting) like follows:

  image video6791: a man is talking about something evaluating validation performance... 1/71 (6.718832) image video6935: a man is talking to a camera for a video game evaluating validation performance... 2/71 (4.730399) image video6697: a man in a black shirt is playing tennis evaluating validation performance... 3/71 (3.754927) image video6929: a man is running on the field evaluating validation performance... 4/71 (4.739570) image video6629: a person is playing a video game evaluating validation performance... 5/71 (3.456113) ......

  Then you can evaluate the final results with

  python eval_category_wise.py --res_file your_res_file_path

2. Visual features + C3D_16:

Set options -input_feature_dim 15360, -input_h5 data_lexical_C3D_16.h5, -input_json data_lexical_C3D_16.json and follow steps above.

3. Visual features + C3D_16 + C3D_2:

Set options -input_feature_dim 19456, -input_h5 data_lexical_C3D.h5, -input_json data_lexical_C3D.json and follow steps above.

4. Visual features + C3D_16 + C3D_2 + BoAW:

Set options -input_feature_dim 19556, -input_h5 data_lexical_C3D_audio.h5, -input_json data_lexical_C3D_audio.json and follow steps above.

Updated results

Performance on the validation set:

Method	BLEU@4	METEOR	CIDEr	ROUGE-L
Category-wise	40.9	28.2	44.7	61.8
+C3D_16	42.2	28.7	46.8	61.9
+C3D_2	43.4	29.4	49.6	62.8
+audio (BoAW)	44.2	29.4	50.5	62.6

Two improvements for NeuralTalk2

To further improve the language model performance, we modified the vanilla NeuralTalk2 with two aspects.

1. A trick to overcome the GPU memory constrain by accumulating gradients over two training iterations. Set option -update_iter to a larger number if necessary.

2. See issue 87. Following the explanation there, we also replaced the log_probs (p) with log_perplexity (ppl) in the beam search operation. This is more consistent with the optimization function during training, and could give higher BLEU, METEOR, CIDEr and ROUGE-L scores (we used in our updated models).

If you find this helps your research, please consider citing:

 @inproceedings{shen2017weakly,
      title={Weakly Supervised Dense Video Captioning},
      author={Shen, Zhiqiang and Li, Jianguo and Su, Zhou and Li, Minjun and Chen, Yurong and Jiang, Yu-Gang and Xue, Xiangyang},
      booktitle ={CVPR},
      year={2017}
       }

Contact

Zhiqiang Shen (zhiqiangshen0214 at gmail.com)

Any discussions and suggestions are welcome!