by Bang An*, Sicheng Zhu* , Michael-Andrei Panaitescu-Liess , Chaithanya Kumar Mummadi , Furong Huang

About

Vision-language models like CLIP are widely used in zero-shot image classification due to their ability to understand various visual concepts and natural language descriptions. However, how to fully leverage CLIP's unprecedented human-like understanding capabilities to achieve better performance is still an open question. This paper draws inspiration from the human visual perception process: when classifying an object, humans first infer contextual attributes (e.g., background and orientation) which help separate the foreground object from the background, and then classify the object based on this information. Inspired by it, we observe that providing CLIP with contextual attributes improves zero-shot image classification and mitigates reliance on spurious features. We also observe that CLIP itself can reasonably infer the attributes from an image. With these observations, we propose a training-free, two-step zero-shot classification method PerceptionCLIP. Given an image, it first infers contextual attributes (e.g., background) and then performs object classification conditioning on them. Our experiments show that PerceptionCLIP achieves better generalization, group robustness, and interpretability. For example, PerceptionCLIP with ViT-L/14 improves the worst group accuracy by 16.5% on the Waterbirds dataset and by 3.5% on CelebA.

Setup

Setting up conda env (Optional)

conda create -n perceptionclip python=3.10
conda activate perceptionclip
pip install open_clip_torch
pip install git+https://github.com/modestyachts/ImageNetV2_pytorch
pip install kornia
export PYTHONPATH="$PYTHONPATH:$PWD"

Datasets

Create dictionary ./datasets/data and link all datasets to ./datasets/data. Please refer to DATA.md for the structure of dataset dictionaries. Here's an example of how to create a symbolic link:

ln -s PATH_TO_YOUR_DATASET PATH_TO_YOUR_PROJ/datasets/data

Code structure

Here's a brief intro of the major components of the code:

./src/datasets contains the code for all the Datasets and Dataloaders.
./src/templates contains all the text prompts.
./src/zero_shot_inference contains the major code for our method and experiments.
./scripts contains all the running scripts.
./replicate_runs.sh This script calls other scripts in the ./scripts directory with the necessary parameters to replicate our experiments.
./visualizations contains the code for visualizations.

To replicate all our experiments, please refer to ./replicate_runs.sh. Below, we provide a few examples demonstrating how to run the code using commands.

CLIP understands contextual attributes

Followings are example commands to reproduce Figure 3.

# evaluate similarity score w/o z
python src/zero_shot_inference/eval_similarity.py --model=ViT-B/16 --eval_augmentation=vflip --template=simple_template  --save_name=sim_imagenet_vit16_wo
# evaluate similarity score w/ z_correct
python src/zero_shot_inference/eval_similarity.py --model=ViT-B/16 --eval_augmentation=vflip --template=vflip_template  --save_name=sim_imagenet_vit16_correct
# evaluate similarity score w/ z_wrong
python src/zero_shot_inference/eval_similarity.py --model=ViT-B/16 --eval_augmentation=vflip --template=vflip_template_wrong  --save_name=sim_imagenet_vit16_wrong
# evaluate similarity score w/ z_random
python src/zero_shot_inference/eval_similarity.py --model=ViT-B/16 --eval_augmentation=vflip --template=vflip_template  --random_descrip=True --save_name=sim_imagenet_vit16_random

CLIP benefits from contextual attributes

Followings are example commands to reproduce Table 2.

# evaluate acc w/o z
python src/zero_shot_inference/zero_shot_org.py --model=ViT-B/16 --dataset=ImageNet --eval_augmentation=vflip --template=simple_template --save_name=acc_imagenet_vit16_wo
# evaluate acc w/ z_correct
python src/zero_shot_inference/zero_shot_org.py --model=ViT-B/16 --dataset=ImageNet --eval_augmentation=vflip --template=vflip_template --save_name=acc_imagenet_vit16_correct
# evaluate acc w/ z_wrong
python src/zero_shot_inference/zero_shot_org.py --model=ViT-B/16 --dataset=ImageNet --eval_augmentation=vflip --template=vflip_template_wrong --save_name=acc_imagenet_vit16_vflip_template_wrong
# evaluate acc w/ z_random
python src/zero_shot_inference/zero_shot_org.py --model=ViT-B/16 --dataset=ImageNet --eval_augmentation=vflip --template=vflip_template --random_descrip=True --save_name=acc_imagenet_vit16_random
# evaluate acc w/ self-inferred z
python src/zero_shot_inference/eval_acc_self_infer.py --model=ViT-B/16 --dataset=ImageNet --eval_augmentation=vflip --template0=vflip_template_wrong --template1=vflip_template --infer_mode=0 --save_name=acc_imagenet_vit16_self_infer_wy

CLIP can infer contextual attributes

Followings are example commands to reproduce Table 3.

# method 1: w/ y
python src/zero_shot_inference/eval_infer_z.py --model=ViT-B/16 --dataset=ImageNet --eval_augmentation=vflip --template0=vflip_template_wrong --template1=vflip_template --infer_mode=0  --save_name=infer_z_imagenet_vit16_wy
# method 1: w/o y
python src/zero_shot_inference/eval_infer_z.py --model=ViT-B/16 --dataset=ImageNet --eval_augmentation=vflip --template0=vflip_template_wrong --template1=vflip_template --infer_mode=1  --save_name=infer_z_imagenet_vit16_woy

PerceptionCLIP improves zero-shot generalization

Followings are example commands to reproduce Table 4.

# consider single attributes
python src/zero_shot_inference/perceptionclip_two_step.py --model=ViT-B/16 --dataset=ImageNet --main_template=imagenet_main_template --factor_templates=imagenet_factor_templates --factors=orientation --infer_mode=0 --temperature=3 --save_name=imagnet_ours_wy_vit16
# consider a composition of multiple attributes
python src/zero_shot_inference/perceptionclip_two_step.py --model=ViT-B/16 --dataset=ImageNet --main_template=imagenet_main_template --factor_templates=imagenet_factor_templates --factors=condition,quality --infer_mode=0 --temperature=3 --save_name=imagnet_ours_wy_vit16

Followings are example commands to reproduce Table 5.

# simple template
python src/zero_shot_inference/perceptionclip_one_step.py  --model=ViT-B/16 --dataset=CUB200  --template=simple_template --save_name=cub200_simple_vit16
# domain template
python src/zero_shot_inference/perceptionclip_one_step.py  --model=ViT-B/16 --dataset=CUB200  --template=cub200_simple_template --save_name=cub200_simple_vit16
# domain template + contextual attributes
python src/zero_shot_inference/perceptionclip_two_step.py --model=ViT-B/16 --dataset=CUB200 --main_template=cub200_main_template --factor_templates=cub200_factor_templates --factors=size,background,condition --convert_text=bird --infer_mode=0 --temperature=1 --save_name=cub200_ours_wy_vit16

PerceptionCLIP improves group robustness

Followings are example commands to reproduce Table 7.

# w/ simple background
python src/zero_shot_inference/perceptionclip_two_step.py --model=RN50 --dataset=Waterbirds --template=waterbirds_background_template --infer_mode=0 --temperature=1 --eval_group=True --eval_trainset=True --save_name=waterbirds_ours_wy_RN50
# w/ complex background
python src/zero_shot_inference/perceptionclip_two_step.py --model=RN50 --dataset=Waterbirds --main_template=waterbirds_main_template --factor_templates=waterbirds_factor_templates --factors=background --infer_mode=0 --temperature=1 --eval_group=True --eval_trainset=True --save_name=waterbirds_ours_wy_RN50_factor

Visualizations

We use Grad-CAM for visualizations. Specifically, we consider the softmax output of the target class and do backpropagation until the targeted layer. Note that this is different from using attentions for visualizations. We use such visualizations to see if the model relies more on the core features when the description of the spurious ones is given while doing classification. Also, this way, adding a spurious attribute’s description won’t really make the model look at it when classifying because all the descriptions (for all classes) will contain that attribute. Please see code and details in ./visualization.

Acknowledgements

This project is based on the following open-source projects. We thank their authors for releasing the source code.

Citing

If you find our work helpful, please cite it with:

@inproceedings{
an2024perceptionclip,
title={Perception{CLIP}: Visual Classification by Inferring and Conditioning on Contexts},
author={Bang An and Sicheng Zhu and Michael-Andrei Panaitescu-Liess and Chaithanya Kumar Mummadi and Furong Huang},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=2Oiee202rd}
}