Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment

Authors: Utkarsh Mall, Cheng Perng Phoo, Meilin Kelsey Liu, Carl Vondrick, Bharath Hariharan, Kavita Bala

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 EXPERIMENTS AND RESULTS
Researcher Affiliation Academia 1Cornell University, Ithaca, NY 2Columbia University, New York, NY
Pseudocode No The paper describes methods in text and equations but does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Our code, data, and other resources are available at: https://graft.cs.cornell.edu
Open Datasets Yes To perform our training, we need a dataset of ground-satellite image pairs. We collected two such datasets for two different kinds of remote sensing imagery: NAIP (U.S.G.S., 2022) (high resolution, with 1 meter per pixel) and Sentinel-2 (Drusch et al., 2012) (low resolution, with 10 meters per pixel). Our dataset collection efforts yield 10.2 million pairs for NAIP and 8.7 million pairs for Sentinel-2 (also refer to Appendix A).
Dataset Splits Yes We select hyperparameters using a validation set with NAIP resolution that we collected. This validation set contains 14 categories with a total of 2632 single-label images.
Hardware Specification No The paper does not specify any hardware details such as GPU models, CPU types, or memory used for running the experiments.
Software Dependencies No The paper mentions software components like Vi T-B/16 and Adam W, but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We train all models for 10 epochs using Adam W with weight decay set to 1e-2. For image-level model, we linearly ramp up the learning rate from 0 to 1e-5 and then decrease the learning rate using a cosine schedule. For pixel-level model, we linearly ramp up the learning rate from 0 to to 5e-5 and then decrease the learning rate to zero using a cosine schedule. All models are initialized using CLIP s weights and the temperature hyperparameter is set to τ = 0.07.