Zoom is what you need: An empirical study of the power of zoom and spatial biases in image classification

Image classifiers are information-discarding machines, by design. Yet, how these models discard information remains mysterious. We hypothesize that one way for image classifiers to reach high accuracy is to first learn to zoom to the most discriminative region in the image and then extract features from there to predict image labels. We study six popular networks ranging from AlexNet to CLIP, and we show that proper framing of the input image can lead to the correct classification of 98.91% of ImageNet images. Furthermore, we explore the potential and limits of zoom transforms in image classification and uncover positional biases in various datasets, especially a strong center bias in two popular datasets: ImageNet-A and ObjectNet. Finally, leveraging our insights into the potential of zoom, we propose a state-of-the-art test-time augmentation (TTA) technique that improves classification accuracy by forcing models to explicitly perform zoom-in operations before making predictions. Our method is more interpretable, accurate, and faster than MEMO, a state-of-the-art TTA method. Additionally, we propose ImageNet-Hard, a new benchmark where zooming in alone often does not help state-of-the-art models better label images.

Our method involves a zooming process that combines resizing and cropping operations. First, we resize the image so that its smaller dimension matches a target scale (S), and then divide it into 3x3 patches. A 224x224px patch is extracted from the center of each patch, with zero padding added if necessary. We demonstrate that this simple approach is sufficient to solve up to 99.44% of the images in the ImageNet dataset.

To understand the potential of using zooming to improve image classification accuracy, first, we establish an upper bound (i.e., when an "optimal" zoom is given). That is, we apply 36 scales * 9 anchors = 324 zoom transformations to each image to generate 324 zoomed versions of the same input. We then feed all 324 versions to each network and label an image "correctly classified given the optimal zoom" if at least 1 of the 324 is correctly labeled. This experiment also importantly informs the community of the type of image that cannot be correctly labeled even with an optimal zooming strategy.

On in-distribution data (ImageNet and ImageNet-ReaL) there exists a substantial improvement when models are provided with an optimal zoom, either selected from 36 (b) or 324 pre-defined zoom crops (c). In contrast, OOD benchmarks still pose a significant challenge to optimal zooming as none of the models (except CLIP) can reach ≥80% accuracy.

State-of-the-art test-time augmentation methods, such as MEMO, finetune a pre-trained classifier at test time to achieve a more accurate prediction. Specifically, MEMO attempts to find a network that produces a low-entropy predicted label over a set of K=16 augmented versions of the test image I and then runs this finetuned model on I again to produce the final prediction. While improving accuracy, MEMO requires a pre-defined set of diverse augmentation transforms (e.g., sheer, rotate, and solarize in AugMix). Yet, the hyperparameters for each type of transform are hard-coded, and the reasons why some of such transform help classification are unknown.

Our goal is to improve MEMO's accuracy and interpretability by replacing all AugMix transforms with only zoom-in transforms. Intuitively, we let a classifier first look at all zoomed-in frames of the input image (at different zoom scales and locations) and then make up its mind to achieve the most confident prediction.

Overall, MEMO + RRC (i.e., random zoom-in transforms) outperforms baselines and the default MEMO.

We present ImageNet-Hard, a novel benchmark in which merely zooming in does not enhance the performance of state-of-the-art models in accurately labeling images. The ImageNet-Hard is a new benchmark that comprises 11,350 images, collected from various existing ImageNet-scale benchmarks (ImageNet, ImageNet-V2, ImageNet-Sketch, ImageNet-C, ImageNet-R, ImageNet-ReaL, ImageNet-A, and ObjectNet). This dataset is challenging to state-of-the-art vision models, as merely zooming in often fails to enhance their ability to classify images correctly. Consequently, even the most advanced models, such as CLIP-ViT-L/14@336px, struggle to perform well on this dataset, achieving only 2.30% accuracy.