Sparse Gaussian Splatting with Fixed Cardinality and Frozen Growth for UAV Scene Reconstruction

Sparse-view UAV imagery provides uneven multi-view supervision for 3D Gaussian Splatting (3DGS), allowing densification to increase model size without comparable reconstruction gains. We propose Structured Gaussian Dropout and Frozen Growth (SGDF) to control active and total primitive counts. Fixed Cardinality (FC) converts the dropout rate into an integer deactivation budget. Frozen Growth (FG) stops later additions and removals while continuous parameters continue to optimize. RISK-A ranks candidate primitives using training-state statistics after protection filtering. Across six scenes with 16, 32, and 48 training views, SGDF achieved mean PSNR, SSIM, and LPIPS values of 25.932 dB, 0.8061, and 0.1804. Relative to 3DGS, PSNR increased by 0.655 dB, while LPIPS decreased by 10.0%. On AeroGrid100-32, final primitive count and full training time decreased by 96.1% and 71.4%, respectively. In the LiDAR evaluation, Full SGDF reduced weighted camera-z RMSE by 15.6% relative to FG. SGDF improves average reconstruction quality across the evaluated UAV scenes while reducing model size and training time on AeroGrid100-32.