maxzanella / mta

[CVPR 2024] Zero-shot method for Vision-Language Models based on a robust formulation of the MeanShift algorithm for Test-time Augmentation (MTA).

On the Test-Time Zero-Shot Generalization of Vision-Language Models: Do We Really Need Prompt Learning? [accepted at CVPR 2024]

Paper: On the Test-Time Zero-Shot Generalization of Vision-Language Models: Do We Really Need Prompt Learning?.

Authors: Maxime Zanella, Ismail Ben Ayed.

This is the official GitHub repository for our paper accepted at CVPR '24. This work introduces the MeanShift Test-time Augmentation (MTA) method, leveraging Vision-Language models without the necessity for prompt learning. Our method randomly augments a single image into N augmented views, then alternates between two key steps (see mta.py and Details on the code section.):

1. Computing a Score for Each Augmented View

This step involves calculating a score for each augmented view to assess its relevance and quality (inlierness score).

Figure 1: Score computation for each augmented view.

2. Seeking for the Mode

Based on the scores computed in the previous step, we seek the mode of the data points (MeanShift).

Figure 2: Seeking the mode, weighted by inlierness scores.

Dataset Preparation and Installation

We follow TPT installation and preprocessing. This ensures that your dataset is formatted appropriately. You can find their repository here. If more convenient you can change the folder names of each dataset in the dictionnary ID_to_DIRNAME in data/datautils.py (line 20).

Quick Start Guide

Running MTA

Execute MTA on the ImageNet dataset with a random seed of 1 and 'a photo of a' prompt by entering the following command:

python main.py --data /path/to/your/data --mta --testsets I --seed 1

Or the 15 datasets at once:

python main.py --data /path/to/your/data --mta --testsets I/A/R/V/K/DTD/Flower102/Food101/Cars/SUN397/Aircraft/Pets/Caltech101/UCF101/eurosat --seed 1

Details on the code

More information on the procedure in mta.py.

Function: gaussian_kernel

• Purpose: Computes the Gaussian kernel density estimate of data points relative to the mode.
• Process:
  • Computes the Euclidean distance between each datapoint and the mode.
  • Applies the Gaussian function to these distances to compute density values.

Function: solve_mta

• Purpose: Solves the MTA alternate procedure.
• Process:
  1. Feature Extraction: Computes image and text features.
  2. Initialization:
     • Calculates pairwise distances among image features to determine the bandwidth for the Gaussian kernel.
     • Initializes the affinity matrix based on the softmax of scaled logits.
     • Sets initial values for inlierness scores (y) uniformly.
  3. Iterative optimization:
     • Alternates between updating inlierness scores and the mode.
  4. Output Calculation: Computes the final output by cosine similarity of the final mode with the text features.

Citation

If you find this project useful, please cite it as follows:

@inproceedings{zanella2024test,
  title={On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning?},
  author={Zanella, Maxime and Ben Ayed, Ismail},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={23783--23793},
  year={2024}
}

Acknowledgement

We express our gratitude to the TPT authors for their open-source contribution. You can find their repository here.

mta's Issues

A question about the second "Text-knowledge guided quadratic term"

Great Paper!
I have one small question though. How does the second term encourages samples with nearby text-based zero-shot predictions to have similar inlierness scores?
When $s_p$ and $s_q$ are alike, $w_{pq}$ is high. Then how are $y_p$ and $y_q$ to be similar?
Could you kindly give me some instructions? Thanks!