Code for SAAT from "Syntax-Aware Action Targeting for Video Captioning" (Accepted to CVPR 2020). The implementation is based on "Consensus-based Sequence Training for Video Captioning".

• Python 3.6
• Pytorch 1.1
• CUDA 10.0
• Microsoft COCO Caption Evaluation
• CIDEr

(Check out the coco-caption and cider projects into your working directory)

Data can be downloaded here (1.6GB). This folder contains:

• input/msrvtt: annotatated captions (note that val_videodatainfo.json is a symbolic link to train_videodatainfo.json)
• output/feature: extracted features of IRv2, C3D and Category embeddings
• output/metadata: preprocessed annotations
• output/model_svo/xe: model file and generated captions on test videos, the reported result can be reproduced by the model provided in this folder (CIDEr 49.1 for XE training)

make -f SpecifiedMakefile test [options]

Please refer to the Makefile (and opts_svo.py file) for the set of available train/test options. For example, to reproduce the reported result

make -f Makefile_msrvtt_svo test GID=0 EXP_NAME=xe FEATS="irv2 c3d category" BFEATS="roi_feat roi_box" USE_RL=0 CST=0 USE_MIXER=0 SCB_CAPTIONS=0 LOGLEVEL=DEBUG LAMBDA=20

To train the model using XE loss

make -f Makefile_msrvtt_svo train GID=0 EXP_NAME=xe FEATS="irv2 c3d category" BFEATS="roi_feat roi_box" USE_RL=0 CST=0 USE_MIXER=0 SCB_CAPTIONS=0 LOGLEVEL=DEBUG MAX_EPOCH=100 LAMBDA=20

If you want to change the input features, modify the FEATS variable in above commands.

@InProceedings{Zheng_2020_CVPR,
author = {Zheng, Qi and Wang, Chaoyue and Tao, Dacheng},
title = {Syntax-Aware Action Targeting for Video Captioning},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

• Pytorch implementation of CST
• PyTorch implementation of SCST