Learning Instance-Level Knowledge With Image-Level Supervision for Open-Vocabulary Object Detection in Remote Sensing Images

Yongbo Li, Yunpeng Bai, Jiaman Ma, Amirkhan A. Temirbayev, Changjing Shang, Qiang Shen

2026

Closed access Metadata (CSL-JSON)
DOI 10.1109/tgrs.2026.3699883
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Open-Vocabulary Object Detection (OVOD) has shown significant potential in real-world remote sensing applications, thanks to its adaptability to open category spaces. Existing OVOD methods mainly rely on pretrained Vision–Language Models (VLMs) to recognize unknown categories. However, VLMs are usually trained on image–text pairs. Whilst possessing strong global semantic modeling capabilities, they struggle to accurately capture the relationship between object instances and their precise locations. To address this problem, the paper presents LILK, which Learns Instance-Level Knowledge from image-level data through weak supervision. Specifically, an image-level weakly supervised knowledge injection module is first created, which represents instance-level semantic information by introducing global classification queries, while aligning image-level supervision with instance-level supervision through a distinct query promotion strategy. Secondly, a quality-aware pseudo-label rectification module is developed, which filters candidate boxes using dual thresholds based on detector confidence and semantic consistency, while incorporating a SAM2-based relocation mechanism to enhance spatial localization accuracy of pseudo labels. Finally, an image-guided query enhancement module is introduced to guide and reinforce detection queries using instance-level image features, providing the detector with stable and modality-consistent visual references. Extensive experimental results on three commonly used remote sensing object detection datasets, DIOR, DOTA, and NWPU VHR-10, demonstrate the efficacy of LILK in performing OVOD tasks.