Answers are grounded in this paper and cite their sources.
Light field (LF) occlusion removal aims to reconstruct scene content occluded by foreground objects by exploiting the multi-view redundancy of sub-aperture images (SAIs). This paper presents a comprehensive comparative study of LF occlusion-removal methods, organized into six methodological families: one covering traditional signal-processing approaches and five deep-learning paradigms, including implicit single-branch reconstruction, explicit dual-branch mask-guided networks with hard binary masks, dual-branch architectures with soft or adaptive confidence masks, representation-based methods using multi-plane images (MPI) or macro-pixel images (MacPI), and hybrid frameworks integrating LFs with vision-language model (VLM) priors. We first establish a unified taxonomy encompassing these architectural families. We then provide consolidated benchmark comparisons under two widely adopted evaluation protocols, analyze reproducibility variations in repeatedly reported baseline results, and investigate the trade-offs among reconstruction accuracy, model complexity, and inference efficiency using parameter count, FLOPs, and latency. Furthermore, we examine the evolution of LF occlusion-removal architectures from the perspectives of masking strategies, scene representations, backbone design, and loss functions. Our analysis indicates that soft confidence-guided masking combined with MPI- or MacPI-based representations achieves the highest self-reported reconstruction quality under the DeOccNet protocol. However, this pattern is not consistent across protocols: under the independently reported ISTY protocol, the top-performing method instead employs an implicit single-branch backbone without explicit masking or an MPI/MacPI representation, suggesting that backbone design may be as influential as masking strategy and representation choice. Lightweight networks with only 11–17M parameters nonetheless achieve accuracy comparable to substantially larger models. In contrast, VLM-guided reconstruction offers a fundamentally different training-free paradigm but remains limited by high inference latency and hallucination artifacts. Finally, we identify benchmark inconsistency, limited generalization to dense LFs, and controllable generative reconstruction as the primary research challenges and promising directions for future work.
No takes yet. Share an insight, caveat, or question.
Senussi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: