TePD: Temporal privileged distillation for amodal counting under structured occlusion

Authors: MANJUNATHA, H - The University Of Texas At Dallas; BORAH, S - University Of Texas At Tyler; SUNDARAVADIVEL, P - University Of Texas At Tyler; Torbert Iii, Henry; TAMIL, L - The University Of Texas At Dallas

Submitted to: Meeting Abstract
Publication Type: Proceedings
Publication Acceptance Date: 5/4/2026
Publication Date: N/A
Citation: N/A

Interpretive Summary: A central challenge in computer vision is the ill-posed problem of occlusion, where the structure of an object must be inferred from only partial observations. While human vision resolves this through temporal continuity and object permanence, modern learning-based models remain constrained to the visible spectrum, relying on synthetic masking or costly amodal annotations to reason hidden geometry. We introduce Temporal Privileged Distillation (TePD), a principled framework that treats irreversible temporal processes as a source of privileged supervision, reframing time not as data variation but as supervisory advantage. We study amodal counting under structured occlusion, where a future, less-occluded observation is available during training but absent at test time, using agricultural chemical defoliation as a natural mechanism that reveals the underlying structure.

Technical Abstract: A central challenge in computer vision is the ill-posed problem of occlusion, where the structure of an object must be inferred from only partial observations. While human vision resolves this through temporal continuity and object permanence, modern learning-based models remain constrained to the visible spectrum, relying on synthetic masking or costly amodal annotations to reason hidden geometry. We introduce Temporal Privileged Distillation (TePD), a principled framework that treats irreversible temporal processes as a source of privileged supervision, reframing time not as data variation but as supervisory advantage. We study amodal counting under structured occlusion, where a future, less-occluded observation is available during training but absent at test time, using agricultural chemical defoliation as a natural mechanism that reveals the underlying structure. The post-defoliation state is modeled as a Privileged Oracle under Vapnik’s Learning Using Privileged Information (LUPI) paradigm, and a dual-stream architecture distills knowledge from a frozen oracle encoder into a lightweight student operating on occluded inputs. Across extensive validation on the Cotton-TePD benchmark, TePD achieves a Latent Reconstruction Fidelity of (TBD), outperforming supervised baselines by (TBD\%) and offering a label-free, physically grounded route to amodal perception in changing environments.