Transformers are increasingly prevalent for multi-view computer vision tasks, where geometric relationships between viewpoints are critical for 3D perception. To leverage these relationships, multi-view transformers must use camera geometry to ground visual tokens in 3D space. In this work, we compare techniques for conditioning transformers on cameras: token-level raymap encodings, attention-level relative pose encodings, and a new relative encoding we propose -- Projective Positional Encoding (PRoPE) -- that captures complete camera frustums, both intrinsics and extrinsics, as a relative positional encoding. Our experiments begin by showing how relative camera conditioning improves performance in feedforward novel view synthesis, with further gains from PRoPE. This holds across settings: scenes with both shared and varying intrinsics, when combining token- and attention-level conditioning, and for generalization to inputs with out-of-distribution sequence lengths and camera intrinsics. We then verify that these benefits persist for different tasks, stereo depth estimation and discriminative spatial cognition, as well as larger model sizes.

Identifying the land cover category for each pixel in a hyperspectral image (HSI) relies on spectral and spatial information. An HSI cuboid with a specific patch size is utilized to extract spatial-spectral feature representation for the central pixel. In this article, we investigate that scene-specific but not essential correlations may be recorded in an HSI cuboid. This additional information improves the model performance on existing HSI datasets and makes it hard to properly evaluate the ability of a model. We refer to this problem as the spatial overfitting issue and utilize strict experimental settings to avoid it. We further propose a multiview transformer for HSI classification, which consists of multiview principal component analysis (MPCA), spectral encoder-decoder (SED), and spatial-pooling tokenization transformer (SPTT). MPCA performs dimension reduction on an HSI via constructing spectral multiview observations and applying PCA on each view data to extract low-dimensional view representation. The combination of view representations, named multiview representation, is the dimension reduction output of the MPCA. To aggregate the multiview information, a fully-convolutional SED with a U-shape in spectral dimension is introduced to extract a multiview feature map. SPTT transforms the multiview features into tokens using the spatial-pooling tokenization strategy and learns robust and discriminative spatial-spectral features for land cover identification. Classification is conducted with a linear classifier. Experiments on three HSI datasets with rigid settings demonstrate the superiority of the proposed multiview transformer over the state-of-the-art methods.