Generalizable Occlusion-aware Human Novel View Synthesis from Image Pairs

We propose a novel approach for human novel view synthesis from image pairs, achieving high-quality cross-scene rendering. Existing methods struggle with occlusions, often suffering from excessive invalid sampling and sensitivity to depth estimation errors. To address these challenges, we propose GoRF, an innovative framework built upon Neural Radiance Fields (NeRF). Specifically, we develop a geometry-guided occlusion-aware mechanism that implicitly models cross-view geometric projection discrepancies and dynamically adjusts multi-view feature blending weights, mitigating occlusion ambiguity between input views. Furthermore, to enhance detail reconstruction, we propose a hybrid sampling strategy that integrates surface-guided and global sampling, effectively compensating for inaccuracies in depth estimation. By combining these strategies, our method enables occlusion-aware, high-fidelity human novel view synthesis. Extensive experiments on diverse human datasets, including THuman2.0, THuman-Sit, and DNA-Rendering, demonstrate that our approach outperforms state-of-the-art methods in both cross-pose and cross-identity scenarios.