This repo contains video captioning using deep learning implementation using PyTorch of :

• S2VT model as described in this paper.
• SCN-LSTM with Scheduled Sampling as described in this paper

Detailed explanation can be found here.

To install all depedencies, I suggest using miniconda to configure virtual environment easily. After installing miniconda, create the environment using env file according to your operating system by typing:

//on windows
conda env create -f env-windows.yml

//on linux
conda env create -f env-linux.yml

Download the annotation file from here.

Download the extracted MSVD dataset from here. It contains 3 directory which are train, testing, and validation. Each folder contains .npy files, each contains CNN features extracted using VGG for a video.

Download the extracted MSVD dataset from here. It contains CNN features extracted from ECO and ResNext architecture, and also semantics features as described in the original paper. Extract the files msvd_resnext_eco.npy, msvd_semantic_tag_e1000.npy, msvd_tag_gt_4_msvd.npy and put all of them into a single directory.

To train from scratch, use the following command:

python train.py \
--annotation-path <PATH_TO_ANNOTATION> \
--train-data-dir <PATH_TO_TRAIN_DIR> \
--val-data-dir <PATH_TO_VAL_DIR> \
--batch-size 10 --epoch 800 --learning-rate 1e-4 --ckpt-interval 10

To train from scratch, use the following command:

python savc_train.py \
--annotation-path <PATH_TO_ANNOTATION>\
 --dataset-dir <PATH_TO_DATASET> \
 --batch-size 64 --epoch 150 --learning-rate 5e-4 --ckpt-interval 10 \
 --model-cnn-2d "resnext" --model-cnn-3d "eco" \
 --mode sample --timestep 30

python generate_caption.py \
--annotation-path <PATH_TO_ANNOTATION> \
--test-data-dir <PATH_TO_TEST_DIR> \
--model-path <PRETRAINED_MODEL_PATH> \
--batch-size 10 \
--out-path <OUT_PATH>

python savc_generate_caption.py \
--annotation-path <PATH_TO_ANNOTATION> \
--dataset-dir <PATH_TO_DATASET> \
--model-path <PRETRAINED_MODEL_PATH> \
--batch-size 64 --mode argmax --out-path <OUT_PATH>

To get the score metrics on Blue, CiDEr, ROUGE_L, and METEOR, generate the captions and use this repository.

CIDEr: 0.6043070425344274
Bleu_4: 0.3966372294884941
Bleu_3: 0.5054434967645823
Bleu_2: 0.6120306365847461
Bleu_1: 0.7489382962776046
ROUGE_L: 0.6659791835245382
METEOR: 0.3006579474917490

CIDEr: 1.1695504045547098
Bleu_4: 0.6404037901180769
Bleu_3: 0.7334661777729088
Bleu_2: 0.8244963341487743
Bleu_1: 0.9204766536962741
ROUGE_L: 0.7941562007691393
METEOR: 0.4155076768981118

• Venugopalan, S., Rohrbach, M., Donahue, J., Mooney, R., Darrell, T., & Saenko, K. (2015). Sequence to sequence - Video to text. Proceedings of the IEEE International Conference on Computer Vision, 2015 Inter, 4534–4542. https://doi.org/10.1109/ICCV.2015.515
• Chen, H., Lin, K., Maye, A., Li, J., & Hu, X. (2020). A Semantics-Assisted Video Captioning Model Trained With Scheduled Sampling. Frontiers in Robotics and AI, 7, 1–11. https://doi.org/10.3389/frobt.2020.475767
• https://github.com/WingsBrokenAngel/Semantics-AssistedVideoCaptioning
• https://github.com/mzolfaghari/ECO-pytorch
• https://github.com/okankop/Efficient-3DCNNs
• https://github.com/YiyongHuang/S2VT