Next Steps

As we continue to refine our pipeline, several enhancements are planned to improve robustness, usability, and reconstruction quality over the upcoming fall semester:

SLAM Optimization

We aim to fine-tune key SLAM parameters to improve both trajectory accuracy and map consistency. Tuning these parameters for 360° panoramic video, particularly in outdoor environments with dynamic lighting and low-texture regions, will help reduce drift and produce cleaner initial reconstructions.

Semantic Filtering

We plan to integrate a semantic segmentation module into the preprocessing stage. This will allow us to identify and mask dynamic objects such as pedestrians, vehicles, and animals during GS scene reconstruction. We aim to use a semantic segmentation model on the panorama images during the SLAM process to create masks for objects to be filtered in the post-processing reconstruction.

Confidence Visualization

We will compute per-frame confidence scores based on factors such as feature density, motion blur, and photometric stability. These scores will be rendered into the SLAM UI that will allow users to visually assess data quality in real-time or during post-capture review.

We are actively looking at various metrics to compute based on both sparse point cloud, dense point cloud and images.

Adaptive Sampling

To optimize efficiency during training and evaluation, we plan to implement adaptive subsampling strategies guided by the confidence maps. Instead of using all frames uniformly, the system will prioritize high-confidence, spatially diverse views, reducing redundancy and computational overhead without sacrificing reconstruction quality.