Self-Improving 4D Perception via Self-Distillation
This addresses scalability issues in 4D perception for computer vision researchers by enabling self-improvement without external annotations, though it is incremental as it builds on existing pretrained models.
The paper tackles the problem of expensive ground-truth annotations for multi-view reconstruction in dynamic scenes by proposing SelfEvo, a self-improving framework that uses unlabeled videos and self-distillation, achieving up to 36.5% improvement in video depth estimation and 20.1% in camera estimation without labeled data.
Large-scale multi-view reconstruction models have made remarkable progress, but most existing approaches still rely on fully supervised training with ground-truth 3D/4D annotations. Such annotations are expensive and particularly scarce for dynamic scenes, limiting scalability. We propose SelfEvo, a self-improving framework that continually improves pretrained multi-view reconstruction models using unlabeled videos. SelfEvo introduces a self-distillation scheme using spatiotemporal context asymmetry, enabling self-improvement for learning-based 4D perception without external annotations. We systematically study design choices that make self-improvement effective, including loss signals, forms of asymmetry, and other training strategies. Across eight benchmarks spanning diverse datasets and domains, SelfEvo consistently improves pretrained baselines and generalizes across base models (e.g. VGGT and $Ï^3$), with significant gains on dynamic scenes. Overall, SelfEvo achieves up to 36.5% relative improvement in video depth estimation and 20.1% in camera estimation, without using any labeled data. Project Page: https://self-evo.github.io/.