DynamicVis: Dynamic visual perception for efficient remote sensing foundation models

The continuous advancement of remote sensing technology has enabled high-resolution Earth observation; however, interpreting these expansive images using modern Vision Foundation Models (VFMs) remains a significant challenge. Unlike object-centric natural images, remote sensing imagery is fundamentally characterized by extreme target sparsity and massive spatial redundancy. Key objects of interest (e.g., ships, vehicles, scattered buildings) often occupy less than 1% of the spatial extent, surrounded by vast, target-free backgrounds. Many existing VFMs rely on uniform dense processing (e.g., Vision Transformers), while a substantial branch adopts pixel-reconstruction pre-training paradigms (e.g., Masked Autoencoders). These designs may waste substantial computational capacity on modeling redundant backgrounds and dilute the feature representations of small, sparse targets. To bridge this structural misalignment, we propose DynamicVis , a remote sensing visual foundation model explicitly tailored to spatial sparsity in high-resolution imagery, with a focus on large-scale visual pre-training and transferable spatial representations for diverse downstream perception tasks. Architecturally, DynamicVis introduces a Dynamic Region-Aware State Space Model (SSM) that bypasses uniform computation. It adaptively routes and incrementally models only task-relevant, high-salience tokens while employing a parameter-free integration for background context, drastically reducing the complexity of processing ultra-long 2D token sequences ( ∼ 100,000). Crucially, to equip the network with robust spatial-selection capabilities, we propose a novel weakly supervised Region-Level Meta-Embedding Multi-Instance Learning (MIL) pre-training paradigm. Trained on a million-scale dataset, this paradigm promotes foreground-aware regional discrimination in the latent semantic space, mitigating the semantic ambiguity of conventional pixel-reconstruction methods. Extensive evaluations across nine diverse downstream visual perception tasks reveal that DynamicVis achieves competitive performance across region-, instance-, and pixel-level remote sensing tasks, with particularly strong results in sparse-target and instance-level perception tasks (e.g., small object detection and instance segmentation). Notably, DynamicVis processes high-resolution imagery (2048 × 2048) with merely 97 ms latency and 833 MB of memory consumption without employing any acceleration techniques (approximately 6% and 3% of a base ViT, respectively), offering a favorable trade-off between computational scalability and representation accuracy for multi-granular remote sensing analysis. Source code is available at https://github.com/KyanChen/DynamicVis .