A comprehensive comparative study has evaluated the performance of various light field (LF) occlusion-removal methods, highlighting significant variations in efficacy across different evaluation protocols. Researchers analyzed six methodological families, including traditional signal-processing techniques and five deep-learning paradigms. These ranged from implicit single-branch reconstruction networks to hybrid frameworks that integrate light field data with vision-language model (VLM) priors, providing a unified taxonomy for comparing these diverse architectural approaches.
The study found that the choice of evaluation protocol significantly influences the perceived success of specific methods. When measured against the DeOccNet protocol, architectures utilizing soft confidence-guided masking in combination with multi-plane image (MPI) or macro-pixel image (MacPI) representations achieved the highest self-reported reconstruction quality. Conversely, when assessed using the independently reported ISTY protocol, the top-performing architecture was an implicit single-branch backbone that did not incorporate explicit masking or MPI/MacPI representations. This suggests that the influence of backbone design is often as significant as masking strategy and scene representation choice.
In terms of computational efficiency, the investigation noted that network size does not always correlate directly with reconstruction accuracy. Lightweight neural networks, containing only 11 to 17 million parameters, were capable of achieving performance levels comparable to substantially larger models. This indicates that more efficient, lower-parameter architectures can meet the requirements for high-quality occlusion removal without the increased model complexity associated with larger systems.
The researchers also evaluated VLM-guided reconstruction, which offers a distinct, training-free paradigm for light field processing. While this approach represents a departure from traditional deep-learning methods, it currently faces significant limitations, specifically regarding high inference latency and the presence of hallucination artifacts.
The analysis identified three primary challenges that continue to impede progress in the field: benchmark inconsistency, limited generalization capabilities when applied to dense light fields, and the need for controllable generative reconstruction. By mapping these technical trade-offs—including parameter counts, floating-point operations (FLOPs), and inference latency—the study provides a baseline for understanding the current landscape of occlusion removal.
Ultimately, the researchers emphasize that there is no universal optimal method for light field occlusion removal, as performance is heavily dependent on both the chosen architecture and the specific evaluation framework used. This comparative work serves to clarify how different design choices—such as masking strategies, backbones, and loss functions—interact to affect overall reconstruction accuracy and efficiency in modern light field processing tasks.
Subject of Research: Light field occlusion removal methodologies
Article Title: Deep learning-based light field occlusion removal: a comparative study
Article References: Senussi, M. F., Abdelbasset, A. G., & Kang, H.-S. (2026). Deep learning-based light field occlusion removal: a comparative study. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11738-9
Image Credits: AI Generated
DOI: 10.1007/s10462-026-11738-9
Keywords: Light field, Occlusion removal, Deep learning, Computer vision, Comparative analysis, Image reconstruction, Neural networks, Benchmark analysis, Multi-plane images, Computational imaging, Inference efficiency, Masking strategies, Backbone design, Model complexity, Vision-language models
News Source: Blake Davidson. (November 17, 2026). Comparative Study Reveals Disparate Performance in Light Field Occlusion Removal Algorithms. Scienmag.



