Frozen Feature Augmentation

Frozen Feature Augmentation
for Few-Shot Image Classification

¹Google DeepMind ²Technische Universität Braunschweig
^*Andreas did the work during an internship at Google DeepMind. ^†Manoj lead the project.

IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2024

Abstract

Training a linear classifier or lightweight model on top of pretrained vision model outputs, so-called 'frozen features', leads to impressive performance on a number of downstream few-shot tasks. Currently, frozen features are not modified during training. On the other hand, when networks are trained directly on images, data augmentation is a standard recipe that improves performance with no substantial overhead. In this paper, we conduct an extensive pilot study on few-shot image classification that explores applying data augmentations in the frozen feature space, dubbed 'frozen feature augmentation (FroFA)', covering twenty augmentations in total. Our study demonstrates that adopting a deceptively simple point-wise FroFAs, such as brightness, can improve few-shot performance consistently across three network architectures, three large pretraining datasets, and eight transfer datasets.

Method Overview

Given a (frozen) pretrained vision transformer, with L transformer blocks (TBs), a multi-head attention pooling (MAP) layer, and a classification layer (CL), we select its L-th Transformer block for caching.

Next, we feed images x to cache (frozen) features f.

Finally, we use the cached features f to train a lightweight model on top. We investigate the effect of frozen feature augmentation (FroFA) applied on cached features in a few-shot setup. We use the same setup for all FroFA experiments .

Investigating 18+2 Frozen Feature Augmentations (FroFAs)

We first investigate eighteen standard image augmentations in a frozen feature setup using an L/16 ViT pretrained on JFT-3B. In these experiments, we cache features on few-shotted ILSVRC-2012 and train a lightweight model on top of augmented frozen features. We identify three strong augmentations: brightness, contrast, and posterize. Results for two additional augmentations are provided in the supplementary.

A Closer Look on Brightness, Contrast, and Posterize FroFAs

We apply brightness, contrast, and posterize in a feature-wise manor (c or c²). Again, we use an L/16 ViT pretrained on JFT-3B, cache features on few-shotted ILSVRC-2012, and train a lightweight model on top of augmented frozen features. We identify brightness c²FroFA as the best working FroFA.

Our Best Working FroFA Yields Strong Performance Across Architectures, Pretraining Datasets, and Few-Shot Datasets

Average top-1 accuracy gains on few-shotted ILSVRC-2012, using frozen features from JFT-3B ViTs. We compare to a weight-decayed MAP (MAP^wd) and L2-regularized linear probe baseline. Our method uses augmented frozen features while the baseline methods use unaltered frozen features. In almost all 5- to 25-shot settings, our method matches or surpasses both baselines.

Average top-1 accuracy gains on few-shotted ILSVRC-2012, using frozen features from ImageNet-21k ViTs. We compare to a weight-decayed MAP (MAP^wd) and L2-regularized linear probe baseline. Our method uses augmented frozen features while the baseline methods use unaltered frozen features. In all 5- to 25-shot settings, our method matches or surpasses both baselines.

Average top-1 accuracy gains across seven few-shot test sets, including, CIFAR100 and SUN397. We train on frozen features from an L/16 model pretrained on JFT-3B or an L/16 image encoder of a WebLI-SigLIP pretrained vision-language model. Our method uses augmented frozen features while the baseline methods use unaltered frozen features. On average, our method shows superior performance across shots for both settings.

BibTeX

@InProceedings{Baer2024, author = {Andreas B\"ar and Neil Houlsby and Mostafa Dehghani and Manoj Kumar}, booktitle = {Proc.\ of CVPR}, title = {{Frozen Feature Augmentation for Few-Shot Image Classification}}, month = jun, year = {2024}, address = {Seattle, WA, USA}, }

Frozen Feature Augmentation for Few-Shot Image Classification