A comparative study evaluated six types of light-field de-occlusion methods and found that the best architecture varies depending on the evaluation protocol. Backbone network design and scene representation also influence the results. Under the DeOccNet protocol, methods combining soft confidence masks with MPI or MacPI representations achieved the highest reconstruction quality, while the best-performing approach under the ISTY protocol was an implicit single-branch backbone network. Lightweight networks with 11 to 17 million parameters can achieve performance comparable to larger models, but VLM-guided reconstruction remains limited by high inference latency and hallucination artifacts.
Source:BIOENGINEER.ORG · Read original
Leave a Reply