Lei Li2β Sishuo Chen1β Xu Sun1β Lu Hou3
[2024-08-08] Results of LLaVA-Next-Video, VILA-1.5 and LongVA are added to the leaderboard.
[2024-07] πππ TempCompass is integrated into LMMs-Eval. See here for usage examples.
[2024-06-11] Result of Reka-core is added to the leaderboard.
[2024-05-25] TempCompass Leaderboard is available on HuggingFace Space π€.
[2024-05-16] πππ TempCompass is accepted at ACL 2024 Findings!
[2024-04-14] Evaluation result of Gemini-1.5-pro, the current SOTA Video LLM, is add.
[2024-03-23] The answer prompt is improved to better guide Video LLMs to follow the desired answer formats. The evaluation code now provides an option to disable the use of ChatGPT.
[2024-03-12] π₯π₯π₯ The evaluation code is released now! Feel free to evaluate your own Video LLMs.
- TempCompass encompasses a diverse set of temporal aspects (left) and task formats (right) to comprehensively evaluate the temporal perception capability of Video LLMs.
-
We construct conflicting videos to prevent the models from taking advantage of single-frame bias and language priors.
-
π€ Can your Video LLM correctly answer the following question for both two videos?
What is happening in the video?
A. A person drops down the pineapple
B. A person pushes forward the pineapple
C. A person rotates the pineapple
D. A person picks up the pineapple
To begin with, clone this repository and install some packages:
git clone https://github.com/llyx97/TempCompass.git
cd TempCompass
pip install -r requirements.txt
1. Task Instructions
The task instructions can be found in questions/
.
Task Instruction Generation Procedure
-
Generate Multi-Choice QA instructions (
question_gen.py
). -
Manually validate quality and rectify.
-
Generate task instructions for Yes/No QA (
question_gen_yes_no.py
), Caption Matching (question_gen_caption_match.py
) and Caption Generation (question_gen_captioning.py
), based on manually rectified Multi-Choice QA instructions. -
Manually validate quality and rectify.
2. Videos
All the processed videos can be downloaded from google drive or huggingface.
As an alternative, you can also download the raw videos and process them yourself
Run the following commands. The videos will be saved to videos/
.
cd utils
python download_video.py # Download raw videos
python process_videos.py # Construct conflicting videos
Note: If you encounter a MoviePy error
when running the processing script, please refer to this issue.
We use Video-LLaVA and Gemini as examples to illustrate how to conduct MLLM inference on our benchmark.
1. Video-LLaVA
Enter run_video_llava
and install the environment as instructed.
Then run the following commands. The prediction results will be saved to predictions/video-llava/<task_type>
.
# select <task_type> from multi-choice, yes_no, caption_matching, captioning
python inference_dataset.py --task_type <task_type>
2. Gemini
The inference script for gemini-1.5-pro is run_gemini.ipynb
. It is recommended to run the script in Google Colab.
After obtaining the MLLM predictions, run the following commands to conduct automatic evaluation. Remember to set your own $OPENAI_API_KEY
in utils/eval_utils.py
.
-
Multi-Choice QA
python eval_multi_choice.py --video_llm video-llava
-
Yes/No QA
python eval_yes_no.py --video_llm video-llava
-
Caption Matching
python eval_caption_matching.py --video_llm video-llava
-
Caption Generation
python eval_captioning.py --video_llm video-llava
Tipπ: Except for Caption Generation, you can set --disable_llm
when running the scripts, which will disable chatgpt-based evaluation (i.e., entirely rely on rule-based evaluation). This is useful when you do not want to use ChatGPT API and your MLLM is good at following the instruction to generate answers of specific format.
The results of each data point will be saved to auto_eval_results/video-llava/<task_type>.json
and the overall results on each temporal aspect will be printed out as follows:
{'action': 76.0, 'direction': 35.2, 'speed': 35.6, 'order': 37.7, 'attribute_change': 41.0, 'avg': 45.6}
{'fine-grained action': 58.8, 'coarse-grained action': 90.3, 'object motion': 36.2, 'camera motion': 32.6, 'absolute speed': 47.6, 'relative speed': 28.0, 'order': 37.7, 'color & light change': 43.6, 'size & shape change': 39.4, 'combined change': 41.7, 'other change': 38.9}
Match Success Rate=100.0
Here we provide an example of how to evaluate LLaVA-Next-Video on TempCompass, using lmms-eval.
1. Clone the repo from LLaVA-Next and setup environments
git clone https://github.com/LLaVA-VL/LLaVA-NeXT
cd LLaVA-NeXT
pip install -e .
2. Run inference and evaluation in a single command
accelerate launch --num_processes 8 --main_process_port 12345 -m lmms_eval \
--model llavavid \
--model_args pretrained=lmms-lab/LLaVA-NeXT-Video-32B-Qwen,conv_template=qwen_1_5,video_decode_backend=decord,max_frames_num=32,mm_spatial_pool_mode=average,mm_newline_position=grid,mm_resampler_location=after \
--tasks tempcompass \
--batch_size 1 \
--log_samples \
--log_samples_suffix llava_vid_32B \
--output_path ./logs/
You can also evaluate the performance on each task (e.g., multi-choice) seperately:
accelerate launch --num_processes 8 --main_process_port 12345 -m lmms_eval \
--model llavavid \
--model_args pretrained=lmms-lab/LLaVA-NeXT-Video-32B-Qwen,conv_template=qwen_1_5,video_decode_backend=decord,max_frames_num=32,mm_spatial_pool_mode=average,mm_newline_position=grid,mm_resampler_location=after \
--tasks tempcompass_multi_choice \
--batch_size 1 \
--log_samples \
--log_samples_suffix llava_vid_32B \
--output_path ./logs/
3. Submit results to TempCompass LeaderBoard
Place the lmms-eval outputs (tempcompass_multi_choice.json
, tempcompass_yes_no.json
, tempcompass_caption_matching.json
and tempcompass_captioning.json
) into the same folder and run this script:
python merge_eval_result.py
Then submit the output file merged_result.json
to the leaderboard.
Note: Currently, the evaluation results calculated by lmms-eval on specific temporal aspects might be incorrect (the average accuracy on each task is correct). To obtain the correct results, you can use this script: acc_lmms_eval.py or submit the result to our leaderboard.
The following figures present results of Video-LLaVA, VideoChat2, SPHINX-v2, Gemini-1.5-pro and the random baseline. Results of more Video LLMs and Image LLMs can be found in our paper and the leaderboard.
![Multi-Choice](./assets/multi-choice.jpg)
![Yes/No](./assets/yes_no.jpg)
![Caption Matching](./assets/caption_matching.jpg)
![Caption Generation](./assets/captioning.jpg)
We update the answer prompt for Multi-Choice QA and Caption Matching, from "Best Option:" to "Please directly give the best option:", which can better encourage MLLMs to directly select an option. As such, we can reduce the reliance on ChatGPT API, if an MLLM is good at following the instruction.
The success rate of rule-based matching is as follows:
Multi-Choice QA
V-LLaVA | SPHINX-v2 | LLaMA-VID | Qwen-VL-Chat | PandaGPT | Valley | |
---|---|---|---|---|---|---|
old prompt | 37.9 | 99.6 | 62.9 | 46.8 | 6.4 | 3.5 |
new prompt | 100 | 100 | 97.0 | 98.5 | 3.9 | 0.4 |
Caption Matching
V-LLaVA | SPHINX-v2 | LLaMA-VID | Qwen-VL-Chat | PandaGPT | Valley | |
---|---|---|---|---|---|---|
old prompt | 76.6 | 89.3 | 44.5 | 91.6 | 30.7 | 11.2 |
new prompt | 99.5 | 99.5 | 68.3 | 96.0 | 22.5 | 3.7 |
- Upload scripts to collect and process videos.
- Upload the code for automatic evaluation.
- Upload the code for task instruction generation.
@article{liu2024tempcompass,
title = {TempCompass: Do Video LLMs Really Understand Videos?},
author = {Yuanxin Liu and Shicheng Li and Yi Liu and Yuxiang Wang and Shuhuai Ren and Lei Li and Sishuo Chen and Xu Sun and Lu Hou},
year = {2024},
journal = {arXiv preprint arXiv: 2403.00476}
}