Answers are grounded in this paper and cite their sources.
High-precision detection of refueling interfaces on special equipment is a prerequisite for robots to carry out intelligent refueling operations. However, the lack of datasets in this field, together with the presence of multiple connectors that are visually similar but functionally different in real refueling environments, makes existing methods difficult to transfer effectively. To this end, we constructed the Refueling Interface Dataset, which contains 10 categories, 946 images and 7659 bounding boxes, with an average class-wise cosine similarity of 0.956, forming a fine-grained benchmark of visually similar, weakly textured connectors. Based on YOLO26, we propose ADCF-YOLO, which improves cross-scale feature fusion in the neck and the training supervision. Adaptive Deformable Cross-scale Fusion compensates for the spatial misalignment introduced by upsampling and convolution through a neighborhood convex combination and fuses the two paths with a zero-initialized residual gate. Sub-pixel Phase-Stable Detection (SPSD) exploits the known and invertible sub-pixel shift to map shifted-view predictions back to the original coordinate system, converting an invariance requirement into dense, unambiguous pixel-wise supervision. Counterfactual Context Decoupling (CCD) applies two oppositely directed constraints—annulus replacement and target erasure—to suppress reliance on background features. On the refueling interface dataset, ADCF-YOLO raises mAP@0.5 from 0.934 to 0.959 and mAP@0.5:0.95 from 0.482 to 0.526 over YOLO26n, the highest of the six YOLO-family detectors at this scale, and yields positive gains on two public datasets. The two Transformer-based reference detectors that score higher do so by 0.002–0.007 on the two mAP metrics, at 7.2–16.4 times the GFLOPs and 11.5–11.9 times the parameters. The engineering value lies in deployment: SPSD and CCD act only during training and leave the inference graph and its cost unchanged, so the accuracy gain carries no additional inference cost.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: