Research map / Detection and segmentation / Open-set and few-shot detection

Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm

NeurIPS 2026 · Detection and segmentation › Open-set and few-shot detection · approach: Domain-adaptive object detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces FDEP, which injects frozen DINOv2 features into encoder-decoder infrared small target detectors via a SAMF fusion module and trains with a shared-parameter CO-ISD scheme so only the lightweight branch runs at inference, plus a new HSE metric.

Problem

Infrared small target (SIRST) detection suffers from weak global semantic modeling, a performance–efficiency trade-off, and a fragmented, threshold-sensitive evaluation system.

Approach

Frozen VFM (DINOv2) features are fused into existing SIRST encoder-decoders through semantic-alignment dual-path modulation (SAMF); a parameter-shared main/lightweight training scheme (CO-ISD) transfers semantics implicitly so inference uses only the original lightweight network.

Key results

• On SIRST3, FDEP applied to ACM raises HSE from 68.05 to 75.81 (with CO-ISD inference time 0.022 s, no added inference cost).
• MSDA-Net + FDEP reaches HSE 91.31 on SIRST3 versus 89.72 for the original MSDA-Net, with unchanged inference time (0.035 s).
• CO-ISD cuts inference time of the FDEP-equipped MSDA-Net by 63.0% (0.100 s to 0.037 s) with only a negligible HSE drop (91.73 to 91.31).

Surprising findings

• Directly applying frozen DINO to SIRST without task-specific structure yields very poor HSE (21.77), showing a large domain gap.
• Adding explicit distillation loss on top of CO-ISD did not help and slightly degraded performance (91.31 vs 90.47 without sharing both E+D with EDL is lower).

Limitations

• Gains on the NUDT-SIRST and IRSTD-1K datasets are smaller and some HSE-T values slightly decline, so improvements are not uniform.
• HSE metric is newly proposed and its stability and advantage over ROC-AUC rely on the authors' analysis.
• Training cost of the main branch with a ViT-L/14 VFM is higher, though removed at inference.

Experimental setup

Datasets: SIRST3 NUAA-SIRST NUDT-SIRST IRSTD-1K

Benchmarks: SIRST3 NUAA-SIRST NUDT-SIRST IRSTD-1K

Metrics: IoU nIoU Pd Fa HSE-P HSE-T HSE

Baselines: ACM ALCNet MLCL-Net ALCL-Net DNA-Net GGL-Net UIU-Net MSDA-Net SCTransNet RPCANet

Builds on: DINOv2 MSDA-Net EEDM loss

Closest papers by meaning

• Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (AAAI 2025 · similarity 0.91)
• Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.89)
• IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.88)
• Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection (AAAI 2025 · similarity 0.88)
• SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · similarity 0.87)
• Motion Prior Knowledge Learning with Homogeneous Language Descriptions for Moving Infrared Small Target Detection (AAAI 2025 · similarity 0.86)