Learning Object Permanence from Natural Revealing Events: Temporal Privileged Distillation for Amodal Counting

Developmental vision is shaped by repeated experiences with objects that change in visibility. Keep their identity and structure the same. We apply this idea to train models for amodal counting when objects are heavily and systematically occluded. Our method, called Temporal Privileged Distillation (TePD) uses pairs of images: one showing an object partially hidden and another taken later after a natural process like defoliation has removed much of the obstruction. During training a fixed teacher model learns from the later view, while a smaller student model learns from the current occluded image. The student matches the teacher’s representation and keeps important semantic cues needed for counting. At test time the student works on its own; no future image or teacher is needed. We tested this on cotton, a dataset of 3,000 drone images paired before and after cotton plant defoliation. TePD reaches a latent cosine similarity of 0.873 and a counting mean absolute error of 4.2, compared with 0.745 and 7.2 for the strongest evaluated baseline. A controlled synthetic task yields 0.972 similarity, suggesting that the effect is not limited to cotton texture. This method isn’t meant to mimic infant cognition learning directly. It’s a way to test a narrower developmental idea: stable representations can emerge from observing how visibility changes over time. These findings suggest that natural processes that reveal objects like wind, growth, and weather can be sources of supervision for efficient visual learning.