Class-Conditioned Evidence Recovery in Frozen Vision Transformers for Remote Sensing Object Detection

Aerial object detectors must resolve small, closely spaced objects while retaining enough context to recognize large structures. RS-Adapter uses an image-level class prior to update intermediate features in a frozen vision transformer. The prior combines a learned classifier, fixed class prototypes, and class-conditional descriptor statistics. Cross-attention links the prior to local features, which are fused across scales and gated before being added to the transformer. The adapted features guide a denoiser that generates box prompts for the detection decoder. On VisDrone, RS-Adapter achieved 39.2% average precision (AP) and 28.4% small-object AP, 3.1 percentage points above DEIM-D-FINE-L on the latter metric. AP50 was 78.3% on DIOR, 84.5% on NWPU VHR-10, and 87.7% on RSOD. RS-Adapter trained 16.8 million of its 103.2 million parameters and used less peak training memory than full tuning in the VisDrone comparison.

Authors

Institutions

Publication Details

Journal
Remote Sensing
Published
2026-11-16
DOI
https://doi.org/10.3390/rs18223929
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Class-Conditioned Evidence Recovery in Frozen Vision Transformers for Remote Sensing Object Detection

Xiaodong Zhao, Peng Lin, Sheng Yang, Haoxuan Xiang
2 links
Remote Sensing
Advanced Neural Network Applications
article

Class-Conditioned Evidence Recovery in Frozen Vision Transformers for Remote Sensing Object Detection

Xiaodong Zhao, Peng Lin, Sheng Yang, Haoxuan Xiang
article en

Abstract

Aerial object detectors must resolve small, closely spaced objects while retaining enough context to recognize large structures. RS-Adapter uses an image-level class prior to update intermediate features in a frozen vision transformer. The prior combines a learned classifier, fixed class prototypes, and class-conditional descriptor statistics. Cross-attention links the prior to local features, which are fused across scales and gated before being added to the transformer. The adapted features guide a denoiser that generates box prompts for the detection decoder. On VisDrone, RS-Adapter achieved 39.2% average precision (AP) and 28.4% small-object AP, 3.1 percentage points above DEIM-D-FINE-L on the latter metric. AP50 was 78.3% on DIOR, 84.5% on NWPU VHR-10, and 87.7% on RSOD. RS-Adapter trained 16.8 million of its 103.2 million parameters and used less peak training memory than full tuning in the VisDrone comparison.

Remote SensingVol. 18(22)
Hangzhou Dianzi University (CN)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon