TriSim

01 — Overview

Overview

Contrastive remote-sensing image–text retrieval can mistakenly treat semantically related samples as negatives. Similarity thresholds alone are fragile because they do not capture the overlap and gaps between visual and textual modalities.

TriSim builds a joint space from image–image, image–text, and text–text similarity, then treats likely false negatives as tail events rather than relying on a single cross-modal threshold.

02 — Contributions

Key Contributions

• 01

  Models three complementary similarity relationships for more reliable false-negative analysis.

• 02

  Uses extreme-value reasoning and complementary tail-selection strategies to detect anomalous candidate negatives.

• 03

  Refines candidate weighting with saliency differences that emphasize discriminative regions and suppress ambiguity.

03 — Method

Method

TriSim constructs a three-dimensional negative-similarity space and selects tail samples using distance from a dense ellipsoidal center together with high-similarity extremes. The selected samples guide triplet-loss optimization, while intra-modal saliency differences produce masks for a learned gain matrix.

04 — Evaluation

Results

Experiments on two remote-sensing image–text retrieval benchmarks show improved robustness to false negatives and stronger retrieval performance than the compared approaches.

05 — Reference

Citation

@InProceedings{Zheng_2026_CVPR,
  author    = {Zheng, Chengyu and Lu, Hanzhang and Nie, Jie and Du, Shan},
  title     = {TriSim: Tri-Dimensional Similarity Modeling with Extreme Value Theory for False-Negative Mitigation in Remote Sensing Image-Text Retrieval},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month     = {June},
  year      = {2026},
  pages     = {23944--23954}
}