resize-is-all-you-need

Isaac Corley1 · Caleb Robinson2 · Rahul Dodhia2 · Juan M. Lavista Ferres2 · Peyman Najafirad (Paul Rad)1

1University of Texas at San Antonio    2Microsoft AI for Good Research Lab

This is the official repository for the paper, "Revisiting pre-trained remote sensing model benchmarks: resizing and normalization matters" presented at the 2024 CVPR PBVS Workshop.

In this paper, we find that simply resizing and normalizing remote sensing imagery correctly provides a significant boost, particlarly when transferring ImageNet pretrained models to the remote sensing domain

Resizing

Remote sensing benchmark datasets, e.g. EuroSAT -- 64 x 64, commonly have small image sizes due to patches being extracted from large satellite tiles. However, we find that recently, evaluation is being performed at these small image sizes while being trained at larger image sizes.

Normalization

Furthermore, we find that many pretrained geospatial foundation models are sensitive to the standard normalization used during inference. Blindly using ImageNet statistics can significantly degrade representation ability

Figure 3. t-SNE plots of EuroSAT test set embeddings extracted using a ResNet50 pretrained on ImageNet with different preprocessing. (left to right: 32 × 32 with normalization, 224 × 224 without normalization, 224 × 224 with normalization)

Extracting Features

We have provided a sample script for extracting features using various models from the paper from your own folder of remote sensing imagery. Please modify the script to your use case (for best performance you will need the mean/std of your dataset). The extracted features will be saved to output_directory/model_features.npy

python embed.py --model resnet50_pretrained_moco --output-dir outputs --root path/to/your/folder --image-size 224 --batch-size 32 --workers 8 --device cuda:0

Cite

If this work inspired you to properly resize and normalize your images in benchmarking please consider citing our paper

@InProceedings{Corley_2024_CVPR,
    author    = {Corley, Isaac and Robinson, Caleb and Dodhia, Rahul and Ferres, Juan M. Lavista and Najafirad, Peyman},
    title     = {Revisiting Pre-trained Remote Sensing Model Benchmarks: Resizing and Normalization Matters},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2024},
    pages     = {3162-3172}
}