Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

BLIP-2 Fine‑Tuning for Vision Navigation Assistance

License: MIT

Table of Contents

Introduction

This repository contains code and scripts for fine‑tuning the BLIP‑2 vision‑language model on a custom navigation‑assistance dataset. Our goal is to assist visually impaired users by generating semantic and directional navigation instructions from scene images.

Features

  • Fine‑tuning with LoRA (PEFT) for efficiency
  • Data augmentation pipeline for improving generalization
  • Training and evaluation scripts based on Hugging Face Transformers and Datasets
  • Weighted Enhanced BERTScore metric combining semantic correctness and directional accuracy

Results

Experiment BERT F1 Enhanced BERTScore
Original BLIP‑2 (no fine‑tuning) 0.63 0.46
LM fine‑tune only (9M params, 0.24%) 0.69 0.51
LM fine‑tune + Augmented Dataset 0.72 0.55

Limitations

  • Path ambiguity remains an issue: multiple routes can lead to the same destination.
  • Creating a large‑scale, diverse navigation dataset is challenging.
  • Navigation based on a single image lacks temporal context.

Future Work

  • Extend from single‑image to video‑based navigation for richer context.
  • Explore fine‑tuning the Q‑Former for end‑to‑end improvement.
  • Incorporate multi‑modal sensor data (e.g., depth, LIDAR).

License

This project is licensed under the MIT License. See the LICENSE file for details.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages