Perspective Transformer Nets

⚠️ Curated Limitations & Negative Results

• The printed camera convention remains unresolved as an implementation specification. With conventional world-to-camera extrinsics, mapping screen coordinates and disparity back to world coordinates normally requires inverse calibration and extrinsics; the paper instead prints its K[R|t] factorization together with p_s proportional to Theta p_t. A different convention may explain this, and no author implementation was inspected. The record preserves the source formulas without silently inverting a matrix, declaring an author-code defect, or certifying a camera convention. The supplement requires a nonzero denominator for its coordinate and disparity normalization; numeric grid ranges, units and boundary handling are not fully certified here.
• Tent interpolation and maximum projection are piecewise differentiable. Tent knots and ties for the maximum require derivative or subgradient conventions; differentiable training does not mean smoothness everywhere. For fixed sampling coordinates, gradients with respect to voxel values pass through the active maximum and its interpolation weights, but this local property is not a convexity or convergence guarantee for the complete learned encoder–decoder. Padding and out-of-grid sampling behavior remain implementation gaps.
• A maximum along disparity is the specified silhouette surrogate, not a complete physical visibility, ray-termination or alpha-compositing model. Binary target volumes motivate the source explanation, but interpolation can yield fractional samples, and binary ground-truth voxels do not prove that decoder predictions are binary or calibrated occupancy probabilities. The evaluation threshold of 0.5 is a reported evaluation convention, not an asserted hard threshold throughout training.
• Projection-only training omits ground-truth volume supervision but still uses silhouettes, supplied camera viewpoints and correspondence among views of the same object. The original setup assumes a clean white background and constant lighting; real-image masks require additional segmentation. Different shapes, including hidden interior differences, can share the same observed silhouettes. Matching these masks does not uniquely determine a 3D shape, and the visual-hull discussion is a hypothesis rather than an identifiability theorem.
• The encoder curriculum and Equation 5 express the desired view-invariant representation. Exact latent equality, global optimization and a universal single-view reconstruction guarantee are not established. The description of encoder pretraining followed by volume training does not by itself certify which encoder parameters are frozen or updated in every implementation. The complete framework has learned encoder and decoder parameters; the supplied-camera projection component alone is described without a learned localization stage.
• The original network, rendered-image conditions, volume resolution, category choices and optimizer settings bound the empirical evidence. The eight-view experiments in the supplement use the chair category; their conclusions do not establish general recovery under arbitrary camera coverage, occlusion, backgrounds or unseen object distributions. No universal accuracy, runtime or memory advantage, training reproduction or new numerical benchmark is claimed.
• The evidence consists of the exact nine-page official NIPS 2016 main paper and the four-page supplementary PDF extracted from its official ZIP. Publication year, PDF build timestamps and actual source-access times are distinct. The same supplementary bytes were encountered through an unrelated Swapout link and excluded from Swapout evidence; the PTN acquisition follows its own official route. ArXiv metadata does not establish preprint byte equality. Author code, training, benchmarks and the cited view-rotation and space-carving methods were not independently reproduced or fully audited.