WSCLoc: Weakly-Supervised Sparse-View Camera Relocalization

• DOI: 10.48550/arxiv.2403.15272
• Publication Date: 2024-03-22
• Authors: Jialu Wang; Kaichen Zhou; Andrew Markham; Niki Trigoni
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition

Despite the advancements in deep learning for camera relocalization tasks, obtaining ground truth pose labels required training process remains a costly endeavor. While current weakly supervised methods excel lightweight label generation, their performance notably declines scenarios with sparse views. In response to this challenge, we introduce WSCLoc, system capable of being customized various learning-based models enhance under weakly-supervised and view conditions. This is realized two stages. initial stage, WSCLoc employs multilayer perceptron-based structure called WFT-NeRF co-optimize image reconstruction quality information. To ensure stable process, incorporate temporal information as input. Furthermore, instead optimizing SE(3), opt $\mathfrak{sim}(3)$ optimization explicitly enforce scale constraint. second pre-trained WFT-Pose. enhanced by Time-Encoding based Random View Synthesis inter-frame geometric constraints that consider pose, depth, RGB We validate our approaches on publicly available datasets, one outdoor indoor. Our experimental results demonstrate solutions achieve superior estimation accuracy sparse-view scenarios, comparable state-of-the-art methods. will make code available.