Skip to content

Latest commit

 

History

History
245 lines (225 loc) · 9.29 KB

README.md

File metadata and controls

245 lines (225 loc) · 9.29 KB
Raw

Democratizing Fine-grained Visual Recognition with Large Language Models

Mingxuan Liu · Subhankar Roy · Wenjing Li · Zhun Zhong · Nicu Sebe · Elisa Ricci

ICLR 2024

OpenReview | Paper | Project Page

TL;DR: We propose Fine-grained Semantic Category Reasoning (FineR) system to address fine-grained visual recognition without needing expert annotations and knowing category names as a-priori. FineR leverages large language models to identify fine-grained image categories by interpreting visual attributes as text. This allows it to reason about subtle differences between species or objects, outperforming current FGVR methods.

Logo

📣 News:

  • [03/01/2024] We released the code along with the intermediate results (in experiments/, including: super-class, attributes, attribute-description pairs, LLM-prompts, LLM raw answers, parsed LLM answers).
  • [01/15/2024] Our work is accepted to ICLR 2024 🌼! Code is coming soon. See you in Vienna this May!

💾 Installation

Requirements:

  • Linux or macOS with Python ≥ 3.9
  • PyTorch ≥ 2.1.0
  • OpenAI API (optional, if you want to discover semantic concepts using LMMs)
  1. Clone this repository adn move to the project working directory:
git clone https://github.com/OatmealLiu/FineR.git
cd FineR
  1. Install the working environment step-by-step:
conda create -n finer python=3.9.16 -y  # create finer conda environment
conda activate finer  # activate the environment and install dependencies

pip install --upgrade pip  # enable PEP 660 support
pip install -e .  # install dependencies
pip install git+https://github.com/openai/CLIP.git  # install CLIP
  1. If you want to discover the fine-grained semantic concepts by yourself via LLMs, please state your OpenAI key as:
export OPENAI_API_KEY="YOUR_OPENAI_API_KEY"

Note: you can add the above line to your ~/.bashrc or ~/.zshrc file as well

📦 Datasets Preparation

FineR is training-free. So we only need to prepare discovery and evaluation datasets, for which we provide the splits used in our paper in Google Drive.

  1. Download: we can download the 5 fine-grained datasets and the Pokemon dataset by clicking:

‍🦤 Bird-200 · 🚙 Car-196 · 🐕 Dog-120 · 🌼 Flower-102 · 🐈 Pet-37 · 👾 Pokemon-10

or via gdown:

# Go to your datasets storage directory
cd YOUR_DATASETS_DOWNLOAD_FOLDER

# Bird-200
gdown 'https://drive.google.com/uc?id=1iKctvfTNjYD1AEEqTIRGEqESXTVAC2Zg'
# Car-196
gdown 'https://drive.google.com/uc?id=1pKssGj5qej5HEHXsWRsvw-oOpWiedS76'
# Dog-120
gdown 'https://drive.google.com/uc?id=1HzVOIlNu5Tat0_fCWBGw8trx1KbqWKlH'
# Flower-102
gdown 'https://drive.google.com/uc?id=1iRvXYM8WhkM7M1-GLpTsJ8YWZUGDaAIE'
# Pet-37
gdown 'https://drive.google.com/uc?id=1kBuSXnzvh32IaYX_CowOmbzr1kQYST1T'
# Pokemon-10
gdown 'https://drive.google.com/uc?id=1F_UzN5TE-RfNnLKxykUl_EuOHKiey5L-'
  1. Organize: we can now extract the downloaded datasets wherever we want and can organize the downloaded datasets via softlinks (ln -s):
# Go to FineR working directory and replace YOUR_DATASETS_DOWNLOAD_FOLDER with your datasets storage directory
cd FineR/datasets
sh link_local_sets.sh YOUR_DATASETS_DOWNLOAD_FOLDER

after which, the directory will look like the following and we are ready to go:

FineR
    └── datasets
          ├── car_196
          ├── CUB_200_2011
          ├── dogs_120
          ├── flowers_102
          ├── pet_37
          └── pokemon

📋 Evaluation

We provide all the intermediate results, including super-category, useful attributes, attribute-description pairs, LLM-prompts, raw LLM replies, and parsed LLM replies from our disovery-->grouping pipeline under experiments. So that we can directly do evaluation here.

For the experiments using 3 images per class for discovery (paper Tab. 1), we can run them all via sh batch_launcher_eval or one-by-one:

# Bird-200
sh scripts_eval/b_pipe.sh
# Car-196
sh scripts_eval/c_pipe.sh
# Dog-120
sh scripts_eval/d_pipe.sh
# Flower-102
sh scripts_eval/f_pipe.sh
# Pet-37
sh scripts_eval/p_pipe.sh
# Pokemon
sh scripts_eval/poke_pipe.sh

For the experiments using random (long-tailed) images per class for discovery (paper Tab. 2), we can run them all via sh batch_launcher_eval_random or one-by-one:

# Bird-200
sh scripts_eval_random/b_pipe.sh
# Car-196
sh scripts_eval_random/c_pipe.sh
# Dog-120
sh scripts_eval_random/d_pipe.sh
# Flower-102
sh scripts_eval_random/f_pipe.sh
# Pet-37
sh scripts_eval_random/p_pipe.sh

In addtion, we also provide intermediate results of using 1 to 10 images per category for discovery in experiments folder which we used for sensitivity analysis. If you want to run experiments with them, you can simply replace the --num_per_category argument with 1 to 10 in the scripts.

⛓️ Full Pipeline

To run the full pipeline to discover semantic concepts from few image as observations for the experiments using 3 images per class for discovery, we can run them all via sh batch_launcher_fullpipe.sh or one-by-one:

# Bird-200
sh scripts_full_pipeline/b_pipe.sh
# Car-196
sh scripts_full_pipeline/c_pipe.sh
# Dog-120
sh scripts_full_pipeline/d_pipe.sh
# Flower-102
sh scripts_full_pipeline/f_pipe.sh
# Pet-37
sh scripts_full_pipeline/p_pipe.sh
# Pokemon
sh scripts_full_pipeline/poke_pipe.sh

To run the full pipeline to discover semantic concepts from few image as observations for the experiments using random (long-tailed) images per class for discovery, we can run them all via sh batch_launcher_fullpipe_random.sh or one-by-one:

# Bird-200
sh scripts_full_pipeline_random/b_pipe.sh
# Car-196
sh scripts_full_pipeline_random/c_pipe.sh
# Dog-120
sh scripts_full_pipeline_random/d_pipe.sh
# Flower-102
sh scripts_full_pipeline_random/f_pipe.sh
# Pet-37
sh scripts_full_pipeline_random/p_pipe.sh
# Pokemon
sh scripts_full_pipeline_random/poke_pipe.sh

Again, to run experiments with different number of image observations, you can simply replace the --num_per_category argument with 1 to 10 in the scripts.

Besides, to identify the super-category of the datasets and acquiring useful attributes for VQA-VLMs to describe the images, we can do the following multiple rounds to get the attributes from 3 images observations:

# To do them all together
sh batch_launcher_IdentifyAndHowto.sh
# Bird-200
sh scripts_IdentifyAndHowto/b_pipe.sh
# Car-196
sh scripts_IdentifyAndHowto/c_pipe.sh
# Dog-120
sh scripts_IdentifyAndHowto/d_pipe.sh
# Flower-102
sh scripts_IdentifyAndHowto/f_pipe.sh
# Pet-37
sh scripts_IdentifyAndHowto/p_pipe.sh
# Pokemon
sh scripts_IdentifyAndHowto/poke_pipe.sh

or random amount of observations:

# To do them all together
sh batch_launcher_IdentifyAndHowto_random.sh
# Bird-200
sh scripts_IdentifyAndHowto_random/b_pipe.sh
# Car-196
sh scripts_IdentifyAndHowto_random/c_pipe.sh
# Dog-120
sh scripts_IdentifyAndHowto_random/d_pipe.sh
# Flower-102
sh scripts_IdentifyAndHowto_random/f_pipe.sh
# Pet-37
sh scripts_IdentifyAndHowto_random/p_pipe.sh

🗻 Citation

Should you find our paper valuable to your work, we would greatly appreciate it if you could cite it:

@inproceedings{
    liu2024democratizing,
    title={Democratizing Fine-grained Visual Recognition with Large Language Models},
    author={Mingxuan Liu and Subhankar Roy and Wenjing Li and Zhun Zhong and Nicu Sebe and Elisa Ricci},
    booktitle={The Twelfth International Conference on Learning Representations},
    year={2024},
    url={https://openreview.net/forum?id=c7DND1iIgb}
}