Fully self-supervised end-to-end thermal point-cloud reconstruction

Develop fully self-supervised end-to-end reconstruction for TAP3D that does not require explicit human-detection supervision.

Background

TAP3D uses physics-guided self-supervised pretraining to reconstruct thermal measurements, but its complete training pipeline still relies on supervised human-detection information. The authors explicitly identify the absence of fully self-supervised end-to-end reconstruction as an unresolved limitation and leave its development for future work.

References

Because TAP3D requires explicit human detection, fully self-supervised end-to-end reconstruction is not yet supported. As Fig.\ref{fig:ssl_ablation} shows, SSL pretraining consistently improves accuracy across label budgets, reducing DCD by 11.93\% on average; fully self-supervised reconstruction remains future work.

— TAP3D: Thermal-Assisted 3D Human Point Clouds  (2610.11241 - Zhang et al., 8 Oct 2026) in Section 6, Subsection “Ablation Study”