Home

Awesome

Towards Diverse Paragraph Captioning for Untrimmed Videos

This repository contains PyTorch implementation of our paper Towards Diverse Paragraph Captioning for Untrimmed Videos (CVPR 2021).

Requirements

Training & Inference

Data preparation

  1. Download the pre-extracted video features of ActivityNet Captions or Charades Captions datasets from BaiduNetdisk (code: he21).
  2. Decompress the downloaded files to the corresponding dataset folder in the ordered_feature/ directory.

Start training

  1. Train our model without reinforcement learning, * can be activitynet or charades.
$ cd driver
$ CUDA_VISIBLE_DEVICES=0 python transformer.py ../results/*/dm.token/model.json ../results/*/dm.token/path.json --is_train

 If you want to train the model with key frames selection, you can perform the following instruction instead.

$ cd driver
$ CUDA_VISIBLE_DEVICES=0 python transformer.py ../results/*/key_frames/model.json ../results/*/key_frames/path.json --is_train --resume_file ../results/*/key_frames/pretrained.th

 It will achieve a slightly worse result with only a half of the video features used at inference phase for faster decoding. You need to download the pretrained.th model at first for the key-frame selection.

  1. Fine-tune the pretrained model in step 1 with reinforcement learning.
$ cd driver
$ CUDA_VISIBLE_DEVICES=0 python transformer.py ../results/*/dm.token.rl/model.json ../results/*/dm.token.rl/path.json --is_train --resume_file ../results/*/dm.token/model/epoch.*.th

Evaluation

The trained checkpoints have been saved at the results/*/folder/model/ directory. After evaluation, the generated captions (corresponding to the name file in the public_split) and evaluating scores will be saved at results/*/folder/pred/tst/.

$ cd driver
$ CUDA_VISIBLE_DEVICES=0 python transformer.py ../results/*/folder/model.json ../results/*/folder/path.json --eval_set tst --resume_file ../results/*/folder/model/epoch.*.th

We also provide the pretrained models for the ActivityNet dataset here and Charades dataset here, which are re-run and achieve similar results with the paper.

Reference

If you find this repo helpful, please consider citing:

@inproceedings{song2021paragraph,
  title={Towards Diverse Paragraph Captioning for Untrimmed Videos},
  author={Song, Yuqing and Chen, Shizhe and Jin, Qin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2021}
}