GNeSF: Generalizable Neural Semantic Fields Supplementary Material

To extract features from these source views, we employ a network with shared weights. Each vertex of the feature volume grids is projected to the image feature maps and obtains its image features by interpolation. In comparison, our method is able to segment accurately for various scenes. In several instances, our method correctly segments objects while Mask2Former produces incorrect results. Mask2Former fails to segment some objects such as the table in the third row. We show more qualitative comparison with NeuralRecon in Figure 1.