// readme

GaussianScanner v10.4 — QNN Hardware-Accelerated Edition

Single image → Depth Anything 3 Small (Qualcomm’s official fp32 model) → depth map → 3D point cloud.

For performance optimization source code and validation scope, see PERFORMANCE_AUDIT.md. This round did not adjust the version number or update the backup.

For the latest high-resolution/automatic point size refactor, see HIRES_POINTSIZE_REFACTOR.md: Full 2x mode runs 10 inferences in total. Per-point projection determines the automatic point size, replacing the old tile high-frequency injection and fixed scaling anchor.

Source archive

GaussianScanner-v10.4-source.zip contains only self-written source code, Gradle configuration, utility scripts, and documentation, to facilitate public upload. The following items are not included due to size or licensing reasons. Please supply them before building:

Missing itemHow to supply
app/src/main/assets/da3_small_gpu_fp32.tflite (143.7MB, exceeds GitHub’s 100MB single-file…