CVPR 2023 - ABLE-NeRF: Attention-Based Rendering with Learnable Embeddings for Neural Radiance Field

In this episode we discuss ABLE-NeRF: Attention-Based Rendering with Learnable Embeddings for Neural Radiance Field by Zhe Jun Tang, Tat-Jen Cham, Haiyu Zhao. The paper presents an alternative approach to the Neural Radiance Field (NeRF) method for representing 3D scenes that addresses view-dependent effects such as murky glossy and translucent surfaces. The proposed method, called ABLE-NeRF, uses a self-attention-based framework on volumes along a ray and incorporates Learnable Embeddings to capture view-dependent effects. The results show that ABLE-NeRF significantly reduces blurry glossy surfaces and produces realistic translucent surfaces, surpassing Ref-NeRF in all three image quality metrics.