PET reconstruction algorithms have long relied on sinogram rebinning. However, as detectors grow smaller in a recent wave of cutting-edge scanners, individual sensors no longer accrue hundreds of photons. Instead, most detect a single photon or none at all, effectively turning sinogram data into point-cloud measurements. The highly heterogeneous sensitivity of these scanners is another issue. We approach sinogram rebinning in the face of these challenges with a density-estimation framework that promotes knot sparsity in an underlying spline basis.