Figure 1: Target-support use in FT-FSOD (Yu et al., 2026) and PULSE. (a) FT-FSOD uses support
annotations for target-domain fine-tuning and subsequently scores query predictions without explicitly
consulting the support examples. (b) PULSE retains cached support evidence and uses SSCA to
refine classification scores according to support–query agreement. Box tags denote predicted classes,
and bar lengths illustrate schematic confidence. The dashed candidate illustrates score suppression at
the same location. Test images from DeepFish (Fu et al., 2024) are used for illustration; the support
examples, bounding boxes, and score changes are schematic.