Evaluation Setup<\/h2>\n\n\n\n
Gaussian Splatting (GS) reconstructions are evaluated on a held-out split<\/strong>, with 10% of images reserved for evaluation<\/strong>. All pipelines are trained and evaluated using the same splits to ensure fair comparison.<\/p>\n\n\n\n Reconstruction quality is evaluated using standard image-based metrics:<\/p>\n\n\n\n We initially explored using existing 360\u00b0 datasets such as KITTI-360<\/strong> and Matterport3D<\/strong>. However, neither dataset was suitable for our pipeline:<\/p>\n\n\n\n As a result, we collected our own dataset using an Insta360 X4<\/strong> camera. Videos were recorded at 5.7K and 8K resolutions<\/strong> during outdoor walking trajectories. Frames are temporally sampled from the videos and passed into the pipeline. Based on empirical analysis, we sample one frame per second<\/strong>, aligned with keyframes produced by the baseline Stella-VSLAM system.<\/p>\n\n\n\n The final dataset consists of 8 video walks<\/strong>, each 3\u20138 minutes long<\/strong>, captured both with and without loop closure<\/strong>. This enables evaluation of global bundle adjustment behavior across all three pipelines.<\/p>\n\n\n\n We directly use camera poses output by Stella-VSLAM to train Gaussian Splatting and evaluate reconstruction quality on the held-out test split. The same video sequences are processed using our rig-based SfM pipeline<\/strong>, running bundle adjustment only<\/strong>, without scale-aligned monocular depth estimation. We run the complete system with all preprocessing and optimization components enabled:<\/p>\n\n\n\n Final reconstruction quality is evaluated on the same held-out dataset.<\/p>\n\n\n\n Rig-based SfM methods significantly outperform the Stella-VSLAM baseline in perceptual similarity. The depth-prior pipeline achieves the highest PSNR<\/strong>, with improved stability during the middle stages of GS training (\u224810k iterations). SSIM scores for both rig-based SfM pipelines are nearly identical and substantially higher than those of the Stella-VSLAM baseline, indicating improved structural consistency in the reconstructions.<\/p>\n\n\n\n Rig-based SfM significantly improves Gaussian Splatting reconstruction quality over raw Stella-VSLAM poses, yielding large gains across PSNR, SSIM, and LPIPS. Incorporating depth priors further provides consistent, albeit modest, improvements, particularly in perceptual quality, without introducing metric regressions. These results highlight that pose accuracy is the dominant factor in GS performance, while depth priors act as a stabilizing signal that enhances perceptual fidelity under challenging capture conditions.<\/p>\n\n\n\n <\/p>\n","protected":false},"excerpt":{"rendered":" Dataset We evaluate our method on 8 real-world equirectangular video walks captured using an Insta360 camera mounted on a handheld tripod. These sequences represent natural outdoor motion with varying lighting conditions and scene geometry. Evaluation Setup Gaussian Splatting (GS) reconstructions are evaluated on a held-out split, with 10% of images reserved for evaluation. All pipelines … <\/p>\nPipelines Compared<\/h2>\n\n\n\n
\n
Metrics<\/h2>\n\n\n\n
\n
Final Metrics<\/h2>\n\n\n\n
<\/td> LPIPS<\/td> PSNR<\/td> SSIM<\/td><\/tr> Stella-VSlam<\/td> 0.4844<\/td> 19.79<\/td> 0.6909<\/td><\/tr> Rig-based-SfM<\/td> 0.3948<\/td> 22.48<\/td> 0.7465<\/td><\/tr> Rig-based-SfM-Depth-Priors<\/td> 0.3931<\/td> 22.65<\/td> 0.7469<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\n\n\n\nDetailed Description<\/h1>\n\n\n\n
Dataset Collection<\/h3>\n\n\n\n
\n
Evaluation Protocol<\/h2>\n\n\n\n
Baseline: Stella-VSLAM<\/h3>\n\n\n\n
To account for exposure and lighting variations, image\u2013embedding\u2013based appearance modeling<\/strong> is enabled during GS training.<\/p>\n\n\n\nRig-based SfM (No Depth Priors)<\/h3>\n\n\n\n
Reconstruction quality is evaluated using the same metrics.<\/p>\n\n\n\nFull Pipeline (End-to-End)<\/h3>\n\n\n\n
\n
Results and Analysis<\/h2>\n\n\n\n
LPIPS<\/h3>\n\n\n\n
![\"\"](\"https:\/\/mscvprojects.ri.cmu.edu\/2025team4\/wp-content\/uploads\/sites\/120\/2025\/12\/image-1024x351.png\")
<\/figure>\n\n\n\n
Both rig-based variants achieve very similar performance, with the depth-prior pipeline achieving slightly lower LPIPS values<\/strong>, indicating improved perceptual fidelity.<\/p>\n\n\n\nPSNR<\/h3>\n\n\n\n
![\"\"](\"https:\/\/mscvprojects.ri.cmu.edu\/2025team4\/wp-content\/uploads\/sites\/120\/2025\/12\/image-2-1024x341.png\")
<\/figure>\n\n\n\n
We observe a drop in PSNR after ~15k iterations for both rig-based SfM pipelines, likely due to overfitting to view-dependent effects<\/strong> such as specular highlights and shading changes. This behavior can be mitigated by adjusting learning rates or reducing training iterations.<\/p>\n\n\n\nSSIM<\/h3>\n\n\n\n
![\"\"](\"https:\/\/mscvprojects.ri.cmu.edu\/2025team4\/wp-content\/uploads\/sites\/120\/2025\/12\/image-3-1024x343.png\")
<\/figure>\n\n\n\nAnalysis<\/h2>\n\n\n\n