Authors: Shyamgopal Karthik, Karsten Roth, Massimilano Mancini, Zeynep Akata
This repo extends the great code repository of SEARLE, link here.
Given an image and a target modification (e.g an image of the Eiffel tower and the text "without people and at night-time"), Compositional Image Retrieval (CIR) aims to retrieve the relevant target image in a database. While supervised approaches rely on annotating triplets that is costly (i.e. query image, textual modification, and target image), recent research sidesteps this need by using large-scale vision-language models (VLMs), performing Zero-Shot CIR (ZS-CIR). However, state-of-the-art approaches in ZS-CIR still require training task-specific, customized models over large amounts of image-text pairs. In this work, we propose to tackle CIR in a training-free manner via our Compositional Image Retrieval through Vision-by-Language (CIReVL), a simple, yet human-understandable and scalable pipeline that effectively recombines large-scale VLMs with large language models (LLMs). By captioning the reference image using a pre-trained generative VLM and asking a LLM to recompose the caption based on the textual target modification for subsequent retrieval via e.g. CLIP, we achieve modular language reasoning.
In four ZS-CIR benchmarks, we find competitive, in-part state-of-the-art performance - improving over supervised methods. Moreover, the modularity of CIReVL offers simple scalability without re-training, allowing us to both investigate scaling laws and bottlenecks for ZS-CIR while easily scaling up to in parts more than double of previously reported results. Finally, we show that CIReVL makes CIR human-understandable by composing image and text in a modular fashion in the language domain, thereby making it intervenable, allowing to post-hoc re-align failure cases. Code will be released upon acceptance.
After cloning this repository, install the revelant packages using
conda create -n cirevl -y python=3.8
conda activate cirevl
pip install torch==1.11.0 torchvision==0.12.0 transformers==4.24.0 tqdm termcolor pandas==1.4.2 openai==0.28.0 salesforce-lavis open_clip_torch
pip install git+https://github.com/openai/CLIP.git
Note that to use a BLIP(-2) caption model by default, you need access to GPUs that allow for use of bloat16
(e.g. A100
types).
Download the FashionIQ dataset following the instructions in the official repository. After downloading the dataset, ensure that the folder structure matches the following:
├── FASHIONIQ
│ ├── captions
| | ├── cap.dress.[train | val | test].json
| | ├── cap.toptee.[train | val | test].json
| | ├── cap.shirt.[train | val | test].json
│ ├── image_splits
| | ├── split.dress.[train | val | test].json
| | ├── split.toptee.[train | val | test].json
| | ├── split.shirt.[train | val | test].json
│ ├── images
| | ├── [B00006M009.jpg | B00006M00B.jpg | B00006M6IH.jpg | ...]
Download the CIRR dataset following the instructions in the official repository. After downloading the dataset, ensure that the folder structure matches the following:
├── CIRR
│ ├── train
| | ├── [0 | 1 | 2 | ...]
| | | ├── [train-10108-0-img0.png | train-10108-0-img1.png | ...]
│ ├── dev
| | ├── [dev-0-0-img0.png | dev-0-0-img1.png | ...]
│ ├── test1
| | ├── [test1-0-0-img0.png | test1-0-0-img1.png | ...]
│ ├── cirr
| | ├── captions
| | | ├── cap.rc2.[train | val | test1].json
| | ├── image_splits
| | | ├── split.rc2.[train | val | test1].json
Download the CIRCO dataset following the instructions in the official repository. After downloading the dataset, ensure that the folder structure matches the following:
├── CIRCO
│ ├── annotations
| | ├── [val | test].json
│ ├── COCO2017_unlabeled
| | ├── annotations
| | | ├── image_info_unlabeled2017.json
| | ├── unlabeled2017
| | | ├── [000000243611.jpg | 000000535009.jpg | ...]
Setup the GeneCIS benchmark following the instructions in the official repository. You would need to download images from the MS-COCO 2017 validation set and from the VisualGenome1.2 dataset.
Exemplary runs to compute all relevant evaluation metrics across all four benchmark datasets are provided in example_benchmark_runs/example_benchmark_runs.sh
.
For example, to compute retrieval metrics on the Fashion-IQ Dress subset, simply run:
datapath=/mnt/datasets_r/FASHIONIQ
python src/main.py --dataset fashioniq_dress --split val --dataset-path $datapath --preload img_features captions mods --llm_prompt prompts.structural_modifier_prompt_fashion --clip ViT-B-32
This call to src/main.py
includes the majority of relevant handles:
--dataset [name_of_dataset] #Specific dataset to use, s.a. cirr, circo, fashioniq_dress, fashioniq_shirt (...)
--split [val_or_test] #Compute either validation metrics, or generate a test submission file where needed (cirr, circo).
--dataset-path [path_to_dataset_folder]
--preload [list_of_things_to_save_and_preload_if_available] #One can pass img_features, captions and mods (modified captions). Depending on which is passed, the correspondingly generated img_features, BLIP-captions and LLM-modified captions will be stored. If the script is called again using the same parameters, the saved data is loaded instead - which is much quicker. This is particularly useful when switching different models (such as the llm for different modified captions, or the retrieval model via img_features).
--llm_prompt [prompts.name_of_prompt_str] #LLM prompt to use.
--clip [name_of_openclip_model] #OpenCLIP model to use for crossmodal retrieval.
@article{karthik2024visionbylanguage,
title={Vision-by-Language for Training-Free Compositional Image Retrieval},
author={Shyamgopal Karthik and Karsten Roth and Massimiliano Mancini and Zeynep Akata},
journal={International Conference on Learning Representations (ICLR)},
year={2024}
}