search

TK-21st / DeVAn

[ACL24] DeVAn: Dense Video Annotation for Video-Language Models
https://www.tingkai-liu.org/DeVAn/
3 stars 0 forks source link

readme

DeVAn: Dense Video Annotation for Video-Language Models

This repository contains code and data related to submission to ACL ARR 2024 - DeVAn: Dense Video Annotation for Video-Language Models.

For more details on the dataset and example videos and annotations, refer to our website.

Data

Please see ~data/devan_acl24_release_v1.jsonl.gz~ data/devan_acl24_release_v1.1.jsonl.gz for V1 of release data. V1 release contains 8K videos, their corresponding youtube id, start/end timestamp, ground truth captions/summaries and predicted caption/summaries from models evaluated.

Refer to the data README for instructions of downloading and post-processing videos from Youtube.

Evaluation

Inference

To reproduce inference results reported in the manuscript, please first ensure that you have downloaded the video data following steps listed above.

  1. Video-ChatGPT: Please follow descriptions in official repo, then modify inference/videochatgpt/infer.py based on your input/output/ckpt paths.
  2. ImageBind-LLM: Please follow descriptions in official repo, then modify inference/infer_imagebind_llm.py based on your input/output/ckpt paths.
  3. Video-LLaMA: Please follow descriptions in official repo, then modify inference/infer_videollama.py based on your input/output/ckpt paths.
  4. VideoCoCa: Code and ckpt for our finetuned VideoCoCa model will be released upon publication.

Evaluation

To reproduce the evaluation results reported in the manuscript, please first ensure that you have completed the inference step above.

For generation tasks, refer to evaluation/ngram_metrics.py and evaluation/model_based_metrics.py to compute N-gram based and model-based metrics for model inference results.

Leaderboard

We evaluate a range of different models on both video-to-text generation and text-to-video retrieval tasks.

Caption Tasks

Model Audio BLEU-4 ROUGE-L CIDEr BLEURT R@1 R@5 R@10
Human (Avg) Raw 6.3 32.1 53.9 50.5 - - -
Human (Min) Raw 4.5 29.5 47.1 48.6 - - -
ImageBind-LLM N/A 0.3 20.0 2.1 34.0 - - -
Video-LLaMA2-Instruct 13B N/A 0.1 7.9 0.0 47.2 - - -
Video-LLaMA2-Instruct 13B Raw 0.1 7.9 0.0 47.1 - - -
Video-LLaMA2-Instruct 7B N/A 0.1 10.8 0.0 43.6 - - -
Video-LLaMA2-Instruct 7B Raw 0.1 10.8 0.0 43.6 - - -
VideoChatGPT N/A 0.4 19.9 2.0 40.5 - - -
VideoCoCa N/A 0.2 13.2 2.3 17.6 32% 50% 58%
VideoCoCa ASR 0.8 20.3 9.2 21.9 36% 53% 59%

Summary Tasks

Model Audio BLEU-4 ROUGE-L CIDEr BLEURT R@1 R@5 R@10
Human (Avg) Raw 15.7 34.5 36.9 55.6 - - -
Human (Min) Raw 12.4 32.1 30.9 53.6 - - -
ImageBind-LLM N/A 1.5 22.7 1.1 45.8 - - -
Video-LLaMA2-Instruct 13B N/A 0.5 18.2 0.0 39.9 - - -
Video-LLaMA2-Instruct 13B Raw 0.5 18.2 0.0 40.0 - - -
Video-LLaMA2-Instruct 7B N/A 0.5 19.1 0.0 43.9 - - -
Video-LLaMA2-Instruct 7B Raw 0.5 19.1 0.1 43.9 - - -
VideoChatGPT N/A 2.9 24.4 5.8 46.7 - - -
VideoCoCa N/A 0.9 16.4 3.3 23.9 25% 41% 48%
VideoCoCa ASR 2.0 21.6 5.5 22.9 27% 42% 48%