Awesome

<div align="center"> <h1> EgoPlan-Bench: Benchmarking Multimodal Large Language Models for Human-Level Planning

</h1> <p align="center"> <img src="./figures/intro.png" width="100%" height="100%" /> </p> </div>

📌Table of Contents

Introduction
Data
Getting Started
Submission
Citation
Acknowledgement

🚀Introduction

The pursuit of artificial general intelligence (AGI) has been accelerated by Multimodal Large Language Models (MLLMs), which exhibit superior reasoning, generalization capabilities, and proficiency in processing multimodal inputs. A crucial milestone in the evolution of AGI is the attainment of human-level planning, a fundamental ability for making informed decisions in complex environments, and solving a wide range of real-world problems. Despite the impressive advancements in MLLMs, a question remains: How far are current MLLMs from achieving human-level planning?

To shed light on this question, we introduce EgoPlan-Bench, a comprehensive benchmark to evaluate the planning abilities of MLLMs in real-world scenarios from an egocentric perspective, mirroring human perception. EgoPlan-Bench emphasizes the evaluation of planning capabilities of MLLMs, featuring realistic tasks, diverse action plans, and intricate visual observations. Our rigorous evaluation of a wide range of MLLMs reveals that EgoPlan-Bench poses significant challenges, highlighting a substantial scope for improvement in MLLMs to achieve human-level task planning. To facilitate this advancement, we further present EgoPlan-IT, a specialized instruction-tuning dataset that effectively enhances model performance on EgoPlan-Bench.

This repository describes the usage of our evaluation data (EgoPlan-Val and EgoPlan-Test) and instruction-tuning data (EgoPlan-IT), and provides the corresponding codes for evaluating and fine-tuning MLLMs on our benchmark. Welcome to evaluate your models and explore methods to enhance the models' EgoPlan capabilities on our benchmark!

📝Data

Egocentric Videos

The EgoPlan datasets are constructed based on the two existing egocentric video sources: Epic-Kitchens-100 and Ego4D.

Download the RGB frames of Epic-Kitchens-100. The folder structure of the dataset is shown below:

EPIC-KITCHENS
└── P01
    └── rgb_frames
        └── P01_01
            ├── frame_0000000001.jpg
            └── ...

Download the videos of Ego4D. The folder structure of the dataset is shown below:

Ego4D
└──v1_288p
    ├── 000786a7-3f9d-4fe6-bfb3-045b368f7d44.mp4
    └── ...

EgoPlan Evaluation Data

Questions from the human-verified evaluation data are formatted as multiple-choice problems. MLLMs need to select the most reasonable answer from four candidate choices. The primary metric is Accuracy.

We divide the evaluation data into two subsets: EgoPlan-Val (containing 3,355 samples) for validation and EgoPlan-Test (containing 1,584 samples) for test, wherein the ground-truth answers of EgoPlan-Test are kept non-public. Below shows an example from the validation set:

{
    "sample_id": 115,
    "video_source": "EpicKitchens",
    "video_id": "P01_13",
    "task_goal": "store cereal",
    "question": "Considering the progress shown in the video and my current observation in the last frame, what action should I take next in order to store cereal?",
    "choice_a": "put cereal box into cupboard",
    "choice_b": "take cereal bag",
    "choice_c": "open cupboard",
    "choice_d": "put cereal bag into cereal box",
    "golden_choice_idx": "A",
    "answer": "put cereal box into cupboard",
    "current_observation_frame": 760,
    "task_progress_metadata": [
        {
            "narration_text": "take cereal bag",
            "start_frame": 36,
            "stop_frame": 105
        },
        {
            "narration_text": "fold cereal bag",
            "start_frame": 111,
            "stop_frame": 253
        },
        {
            "narration_text": "put cereal bag into cereal box",
            "start_frame": 274,
            "stop_frame": 456
        },
        {
            "narration_text": "close cereal box",
            "start_frame": 457,
            "stop_frame": 606
        },
        {
            "narration_text": "open cupboard",
            "start_frame": 689,
            "stop_frame": 760
        }
    ],  
}

EgoPlan Training Data (EgoPlan-IT)

We provide an automatically constructed instruction-tuning dataset EgoPlan_IT, which contains 50K samples, for fine-tuning the model. Below shows an example from EgoPlan-IT:

{
    "sample_id": 39,
    "video_source": "EpicKitchens",
    "video_id": "P07_113",
    "task_goal": "Cut and peel the onion",
    "question": "Considering the progress shown in the video and my current observation in the last frame, what action should I take next in order to cut and peel the onion?",
    "answer": "grab onion",
    "current_observation_frame": 9308,
    "task_progress_metadata": [
        {
            "narration_text": "open drawer",
            "start_frame": 9162,
            "stop_frame": 9203
        },
        {
            "narration_text": "grab knife",
            "start_frame": 9214,
            "stop_frame": 9273
        },
        {
            "narration_text": "close drawer",
            "start_frame": 9272,
            "stop_frame": 9303
        }
    ],
    "negative_answers": [
        "open drawer",
        "grab knife",
        "close drawer",
        "slice onion",
        "remove peel from onion",
        "peel onion"
    ]
}

💻Getting Started

1. Installation

Clone the repo and install dependent packages:

git clone https://github.com/ChenYi99/EgoPlan.git
cd EgoPlan
pip install -r requirements.txt

2. Data Preparation

Prepare Egocentric videos: Download the RGB frames of Epic-Kitchens-100 and the videos of Ego4D.

Prepare EgoPlan datasets: Download the validation data set EgoPlan_validation.json and the training dataset EgoPlan_IT.json. Put these two JSON files under the directory data/.

For details of the data structure, please refer to Data.

3. Model Weights

We use Video-LLaMA as an example for evaluation and instruction-tuning.

Prepare the pretrained model checkpoints

The checkpoint of the vanilla Video-LLaMA can be downloaded from Video-LLaMA-2-7B-Finetuned.
Alternatively, the checkpoint of the Video-LLaMA that has been further tuned on our EgoPlan-IT can be downloaded from EgoPlan-Video-LLaMA-2-7B.

Prepare the pretrained LLM weights

Video-LLaMA is based on Llama2 Chat 7B. The corresponding LLM weights can be downloaded from Llama-2-7b-chat-hf.

Prepare weights for initializing the Visual Encoder and Q-Former (optional)

If the server cannot access the Internet, the following weights should be downloaded in advance:<br>

VIT (eva_vit_g.pth)<br>
Q-Former (blip2_pretrained_flant5xxl.pth)<br>
Bert (bert-base-uncased)

4. Evaluation

Evaluating the Vanilla Video-LLaMA

Config the paths for model weights in video_llama_eval_only_vl.yaml.<br> Set the paths for the project root, Epic-Kitchens-100 RGB frames and Ego4D videos in eval_video_llama.sh.<br> Then, run the script on 1xV100 (32G) GPU:

bash scripts/eval_video_llama.sh

Evaluating the Video-LLaMA Tuned on EgoPlan-IT

Config the paths for model weights in egoplan_video_llama_eval_only_vl.yaml.<br> Set the paths for the project root, Epic-Kitchens-100 RGB frames and Ego4D videos in eval_egoplan_video_llama.sh.<br> Then, run the script on 1xV100 (32G) GPU:

bash scripts/eval_egoplan_video_llama.sh

5. Training

For increasing instruction diversity, in addition to EgoPlan-IT, we also include an additional collection of 164K instruction data following Video-LLaMA:

3K image-based instructions from MiniGPT-4 [link].
150K image-based instructions from LLaVA [link]. The images can be downloaded from here.
11K video-based instructions from VideoChat [link]. The videos can be downloaded following the instructions from the official Github repo of Webvid.

Config the paths for model weights and datasets in visionbranch_stage3_finetune_on_EgoPlan_IT.yaml.<br> Set the path for the project root in finetune_egoplan_video_llama.sh.<br> Then, run the script on 8xV100 (32G) GPUs:

bash scripts/finetune_egoplan_video_llama.sh

🖊️Submission

We are consistently maintaining an EgoPlan-Bench Leaderboard. To show your model's performance on our leaderboard, please contact yichennlp@gmail.com with attached prediction files for the validation and test sets.

We ONLY accept ".json" files. The submitted data format should be like:

[
    {  
        "sample_id": "int",  
        "label": "str"
    },
    ...
]

where the "sample_id" field should be an integer and the "label" field should be a string within ["A","B","C","D"]. An example submission file for the validation set can be found here.

📚Citation

If you find our project helpful, hope you can star our repository and cite our paper as follows:

@article{chen2023egoplan,
  title={EgoPlan-Bench: Benchmarking Multimodal Large Language Models for Human-Level Planning},
  author={Chen, Yi and Ge, Yuying and Ge, Yixiao and Ding, Mingyu and Li, Bohao and Wang, Rui and Xu, Ruifeng and Shan, Ying and Liu, Xihui},
  journal={arXiv preprint arXiv:2312.06722},
  year={2023}
}

🙌Acknowledgement

This repo benefits from Epic-Kitchens, Ego4D, Video-LLaMA, LLaMA, MiniGPT-4, LLaVA, VideoChat. Thanks for their wonderful works!