This week, Team VOXEL advanced two core parts of the O2A pipeline.

On segmentation, the team tested zero-shot approaches (MobileSAM and SAM2) on single-object sequences. Early findings showed that presenting multiple candidate masks and letting the user confirm with a single click improves reliability.

On reconstruction, the team ran an initial COLMAP → 3D Gaussian Splatting workflow from video frames. Results highlighted the need to apply foreground masks early, especially for challenging cases like transparent objects.

The high-level architecture diagram was updated to reflect the current flow from capture and preprocessing through segmentation, mapping, 3D construction, texturing, and Unity export.

Next, the team will continue comparing SAM variants, test 3D reconstruction under varied object/background conditions, explore alternative 3D paths such as Instant-NGP. With the PDR just around the corner, the upcoming weeks promise to bring sharper results, clearer comparisons, and the first polished demos of the O2A pipeline.