Scholarlink

<p>Military camouflage creates two coupled errors for open-vocabulary segmentation: uncertain object boundaries and terrain-induced class confusion. We propose the Military Terrain-Conditioned Decoy-Aware Mask Refinement Network (MT-DAMRNet), a terrain-decoy refinement layer on top of a SAM3/CLIP OVCOS backbone. The proposed design suppresses decoy-like evidence only in ambiguous mask regions and reuses the resulting signal to calibrate class logits with terrain context. On the MHCD2022 benchmark under our open-vocabulary camouflaged object segmentation protocol, MT-DAMRNet achieves the best overall trade-off across both segmentation-only and joint segmentation-classification metrics. In segmentation-only evaluation, it achieves a structural measure of 0.880, an enhanced-alignment measure of 0.915, a weighted F-measure of 0.848, and MAE of 0.040. In joint segmentation-classification evaluation, it reaches a class-aware structural measure of 0.860, a class-aware enhanced-alignment measure of 0.894, a class-aware weighted F-measure of 0.825, cMAE of 0.067, and cIoU of 0.767.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Military terrain-conditioned decoy-aware mask refinement network for open-vocabulary camouflaged object segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Woojin Park,
  • Cheoneum Park,
  • Kampol Woradit,
  • Seongmin Pyo,
  • Haneol Jang

摘要

Military camouflage creates two coupled errors for open-vocabulary segmentation: uncertain object boundaries and terrain-induced class confusion. We propose the Military Terrain-Conditioned Decoy-Aware Mask Refinement Network (MT-DAMRNet), a terrain-decoy refinement layer on top of a SAM3/CLIP OVCOS backbone. The proposed design suppresses decoy-like evidence only in ambiguous mask regions and reuses the resulting signal to calibrate class logits with terrain context. On the MHCD2022 benchmark under our open-vocabulary camouflaged object segmentation protocol, MT-DAMRNet achieves the best overall trade-off across both segmentation-only and joint segmentation-classification metrics. In segmentation-only evaluation, it achieves a structural measure of 0.880, an enhanced-alignment measure of 0.915, a weighted F-measure of 0.848, and MAE of 0.040. In joint segmentation-classification evaluation, it reaches a class-aware structural measure of 0.860, a class-aware enhanced-alignment measure of 0.894, a class-aware weighted F-measure of 0.825, cMAE of 0.067, and cIoU of 0.767.