Dense 3D Reconstruction from Monocular Video via Learning-based SLAM and Gaussian Splatting

Our pipeline integrates advanced transformer-based SLAM with explicit radiance field rendering to achieve high-fidelity reconstruction. The workflow is divided into four core modules:

1. Data Preprocessing & Standardization

We capture 4K monocular videos using smartphones and extract frames at a fixed rate of 10 FPS using FFmpeg. This standardized preprocessing unifies image resolution and file naming conventions, ensuring compatibility across both public benchmarks (7-Scenes) and self-collected SUSTech datasets.

2. Dense Tracking and Mapping (MAST3R-SLAM)

The system employs MAST3R (Multi-view Assistant for 3D Reconstruction) as the SLAM front-end. By leveraging pixel-level transformer matching and global bundle adjustment, it simultaneously estimates accurate camera trajectories (poses.txt) and generates a dense, colored raw point cloud (raw_map.ply) without requiring depth sensors.

3. Standardized Point Cloud Refinement

To bridge the gap between noisy SLAM outputs and high-quality 3DGS training, we implement a self-designed post-processing workflow in CloudCompare:

Statistical Outlier Removal (SOR): Eliminating floating noise caused by motion blur or depth jitter.

Voxel Downsampling: Uniformly reducing point density to optimize training efficiency.

Manual Masking: Filtering irrelevant background outliers to focus on the target scene.

4. 3D Gaussian Splatting & NVS

The refined point cloud serves as a geometric prior to initialize Gaussian ellipsoids. Using the splatfacto framework, the system iteratively optimizes the position, covariance, and color of each Gaussian through photometric loss minimization. The final model supports real-time interactive rendering and smooth novel view synthesis.

By comparing three self-collected scenes, we observe that reconstruction quality is highly dependent on Visual Saliency. The pantry yields the densest point cloud due to rich textures, while the blank corner results in extreme sparsity and floaters. 

We analyze the 3DGS training evolution from 7k to 50k iterations. The visual quality reaches a "sweet spot" at 30k, while 50k iterations achieve photorealistic specular details (PSNR > 44dB). The temporal consistency is verified across different frames of the TUM dataset. 

We evaluate our pipeline across environments with varying texture densities. The Pantry (high texture) yields the most complete geometry due to rich visual features. In contrast, The Blank Wall (texture-less) represents a failure case for monocular SLAM, where lack of salient features leads to sparse points and floating artifacts. 

We identify 7,000 iterations as the visual "Sweet Spot" for small-scale indoor scenes. While PSNR continues to climb up to 44.29 dB at 50k iterations, the model begins to overfit, leading to high-frequency artifacts. The 7k-iteration model maintains a more natural structural integrity.