Ordered Support-Conditioned Query Exchange for Cross-Domain Few-Shot Remote Sensing Detection

Generic proposal selection may overlook the target classes specified by a few-shot support set. We propose support-conditioned query exchange (SCQE) for cross-domain few-shot remote sensing detection using a frozen prompt-free universal region proposal network (PF-RPN). SCQE selects incoming encoder references and their receiving decoder slots, with proposal weights, learned query contents and query count fixed. Each replacement is evaluated by its complete support-conditioned coverage gain, including outgoing coverage loss, under a cumulative perturbation budget on protected native outputs. Vector--Jacobian product sketches estimate interactions among exchanges. The unchanged decoder verifies the complete plan, accepting changed proposals or restoring all cached native PF-RPN outputs before shared recognition. A support-adapted CD-ViTO backend provides class scores while retaining PF-RPN box geometry. SCQE achieves average precision (AP@[.50:.95]) of 24.86, 36.36 and 42.36 on DIOR and 21.47, 27.81 and 34.02 on DOTA with horizontal bounding boxes (DOTA-HBB) for 1-, 5- and 10-shot detection. Gains over native selection range from 7.11 to 8.68 and from 4.93 to 6.27 AP points, respectively. DIOR 5-shot ablations support receiving-slot assignment, complete exchange gain and cumulative perturbation control. Disabling the actual-risk gate reduces AP at an intersection-over-union threshold of 0.75 (AP75) by 2.88 points and yields deployed budget violations in 4.7\% of query--support pairs. In the 5-shot setting, SCQE increases mean query-inference latency over the directly verified prefix-selection baseline by 12.4\% on DIOR and 15.0\% on DOTA-HBB, with the same reported peak allocated GPU memory.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-20
DOI
https://doi.org/10.21203/rs.3.rs-11608186/v1
Primary Topic
Domain Adaptation and Few-Shot Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Ordered Support-Conditioned Query Exchange for Cross-Domain Few-Shot Remote Sensing Detection

Zhipeng Ye, Feng Jiang, Lili Zhou
3 links
Research Square
Domain Adaptation and Few-Shot Learning
preprint

Ordered Support-Conditioned Query Exchange for Cross-Domain Few-Shot Remote Sensing Detection

Zhipeng Ye, Feng Jiang, Lili Zhou
preprint en

Abstract

Generic proposal selection may overlook the target classes specified by a few-shot support set. We propose support-conditioned query exchange (SCQE) for cross-domain few-shot remote sensing detection using a frozen prompt-free universal region proposal network (PF-RPN). SCQE selects incoming encoder references and their receiving decoder slots, with proposal weights, learned query contents and query count fixed. Each replacement is evaluated by its complete support-conditioned coverage gain, including outgoing coverage loss, under a cumulative perturbation budget on protected native outputs. Vector--Jacobian product sketches estimate interactions among exchanges. The unchanged decoder verifies the complete plan, accepting changed proposals or restoring all cached native PF-RPN outputs before shared recognition. A support-adapted CD-ViTO backend provides class scores while retaining PF-RPN box geometry. SCQE achieves average precision (AP@[.50:.95]) of 24.86, 36.36 and 42.36 on DIOR and 21.47, 27.81 and 34.02 on DOTA with horizontal bounding boxes (DOTA-HBB) for 1-, 5- and 10-shot detection. Gains over native selection range from 7.11 to 8.68 and from 4.93 to 6.27 AP points, respectively. DIOR 5-shot ablations support receiving-slot assignment, complete exchange gain and cumulative perturbation control. Disabling the actual-risk gate reduces AP at an intersection-over-union threshold of 0.75 (AP75) by 2.88 points and yields deployed budget violations in 4.7\% of query--support pairs. In the 5-shot setting, SCQE increases mean query-inference latency over the directly verified prefix-selection baseline by 12.4\% on DIOR and 15.0\% on DOTA-HBB, with the same reported peak allocated GPU memory.

Research Square
Taizhou Institute of Science and Technology (CN)
Domain Adaptation and Few-Shot Learning
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
Ordered Support-Conditioned Query Exchange for Cross-Domain Few-Shot Remote Sensing Detection — Zhipeng Ye, Feng Jiang, et al. · Research Square (2026) | TGRS Research Map | TGRS