ADP: Adaptive Discriminative Prototypes for Cross-Domain Few-Shot Object Detection

Cross-domain few-shot object detection (CD-FSOD) presents a fundamental challenge: a detector must recognise novel object categories from as few as one to ten labelled examples per class, frequently in visual domains that differ substantially from the training distribution. Leading methods, such as CD-ViTO, address this by constructing per-class prototype representations through Sinkhorn optimal transport clustering of DINOv2 patch-token features. When categories share fine-grained visual texture, however, the resulting prototypes can overlap in feature space, which degrades nearest-neighbour assignment and reduces detection accuracy. To address this problem, we propose ADP (Adaptive Discriminative Prototypes), a training-free post-processing step that increases inter-class prototype separation through difficulty-conditioned anchored repulsion. ADP computes a per-support-set difficulty score δ to quantify inter-class prototype overlap. When δ is below a low threshold, the original Sinkhorn prototypes are returned unchanged. Otherwise, a repulsion strength and margin proportional to δ are applied through a short anchor-regularised gradient step, keeping the refined prototypes tied to the support evidence. On the NTIRE 2026 splits, every measured δ exceeds that threshold, so this deactivation branch is not exercised and its effect is unverified. On the NTIRE 2026 CD-FSOD benchmark, fixed-strength repulsion has been shown to yield large gains on confused domains while degrading performance on well-separated ones. ADP records a higher mAP than the Sinkhorn baseline in 13 of 18 conditions. The average change is +0.19 mAP at 1-shot, +0.03 at 5-shot, and −0.07 at 10-shot. Each number comes from the challenge’s fixed support file, so these changes are directional evidence from a single deterministic run. Relative to fixed-strength repulsion, ADP avoids that variant’s largest regressions, while the average change itself remains small.