Research code for “Semantic-Geometric Dual Compression: Training-Free Visual Token Reduction for Ultra-High-Resolution Remote Sensing Understanding.”
Results in this repository are transcribed from the manuscript and are not independently reproduced.
DualComp routes a visual-token budget between semantic and geometric streams:
- Router predicts task-adaptive
lambdaandrhocontrols. - SCSA aggregates spatially contiguous semantic regions.
- IGSR retains geometry-sensitive anchors and completes structural paths.
- Fusion concatenates the two streams using lambda-aware scaling.
The repository provides a paper-bounded reference implementation. Parameters omitted by the manuscript remain explicit, and no end-to-end host-model adapter is claimed.
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .Python 3.10 and 3.11 are supported.
import torch
from dualcomp import lambda_weighted_concatenation
semantic = torch.ones(1, 2, 4)
geometric = torch.full((1, 1, 4), 2.0)
output = lambda_weighted_concatenation(semantic, geometric, duality=0.25)
assert output.shape == (1, 3, 4)| Item | Reported value |
|---|---|
| XLRS-Bench average | 53.1% |
| Compression ratio | 42.4x |
| Inference speed | 3.87 s/image |
| Qwen2.5-VL-7B transfer | 47.9% at 10.24x |
These values are copied from the manuscript and are not independent reruns.
Citation metadata is in CITATION.cff. Code in this repository is released under Apache-2.0; external assets remain subject to their own terms.