Figure 2. HeROD pipeline. Text and image features are encoded, and spatial/semantic reasoning priors are derived from the phrase and
image. These priors are injected into reference generation, final prediction, and training loss, guiding the detector toward plausible regions
for more data-efficient learning.