Figure 7.
The framework presents S A X N e R F with three parts. In the M L G Sampling Strategy, rays r originate from a source during circular scanning of an object to obtain projection P. Pixel-level and patch-level sampling with H Hash Encoding produce features F with position and feature inputs. The Lineformer applies L S A B three times with skip connection and concatenation C, followed by fully connected layers f c and X-ray Volume Rendering to estimate I sub pred of r and compare with I sub g t of r using L 2 norm. The Line Segment-based Attention Block includes Layer Normalization, L S M S A, Feed Forward Network, and skip connections. The Line Segment-based Multi-head Self Attention splits X into X sub i, computes Q sub i, K sub i, V sub i using f c, applies matrix multiplication with H sub i and positional embedding E sub i, then grouping and Leaky R e L U to produce Y.Sax-NeRF model (Cai et al., 2024)