Don't Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models

This repository contains the official pytorch implementation of the paper: "Don't Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models".

Updates

• 2024.05.29: Build project page
• 2024.05.29: Paper online
• 2024.05.28: Code release

Observation

Attention bias in LVLMs. Even when the image (V) does not contain information relevant to the query (Q), LVLMs exhibit a tendency for attention to be biased towards a few image tokens (i.e., blind tokens). This phenomenon is observed by averaging the attention weights across all layers when generating the first response token.

Motivation

Impact of blind/non-blind tokens on prediction logits. (Left) Zeroing out image tokens with attention weights higher than the mean + standard deviation, i.e., blind tokens, does not significantly affect the original prediction logits, suggesting that LVLMs may assign high attention weights to tokens that do not carry significant object-discriminative information. Conversely, zeroing out non-blind tokens drastically disrupts the logits, often leading to near 50:50 probabilities, indicating a loss of object-discriminative information. (Right) Similarly, examples demonstrate that zeroing out non-blind tokens results in a loss of discriminative power for previously well-classified instances or produces entirely incorrect predictions, causing a significant drop in performance.

Method: AvisC

Setup

conda create AvisC python=3.10
conda activate AvisC
git clone https://github.com/sangminwoo/AvisC.git
cd AvisC
pip install -r requirements.txt

Models

About model checkpoints preparation

• LLaVA-1.5: Download LLaVA-1.5 merged 7B
• InstructBLIP: Download InstructBLIP

Evaluation

• POPE: bash eval_bench/scripts/pope_eval.sh
  • Need to specify "model", "model_path"
• MME: bash experiments/cd_scripts/mme_eval.sh
  • Need to specify "model", "model_path"
• AMBER: bash experiments/cd_scripts/amber_eval.sh
  • Need to specify "model", "model_path"

About datasets preparation

• Please download and extract the MSCOCO 2014 dataset from this link to your data path for evaluation.
• For MME evaluation, see this link.
• For AMBER evaluation, see this link.

Results

POPE

MME

MME-Fullset

MME-Hallucination

AMBER

LLaVA-Bench Examples

Acknowledgments

This codebase borrows from most notably VCD, OPERA, and LLaVA. Many thanks to the authors for generously sharing their codes!

Citation

If you find this repository helpful for your project, please consider citing our work :)

@article{woo2024dont,
  title={Don't Miss the Forest for the Trees: Attentional Vision Calibration for Large Vision Language Models}, 
  author={Woo, Sangmin and Kim, Donguk and Jang, Jaehyuk and Choi, Yubin and Kim, Changick},
  journal={arXiv preprint arXiv:2405.17820},
  year={2024},
}