Portfolio
Research|2024

SplatSLAM

Real-time 3D mapping and SLAM from monocular RGB video using 3D Gaussian Splatting. Photo-realistic dense reconstruction without depth sensors, built as a Nerfstudio extension.

PythonPyTorchComputer Vision3D ReconstructionCUDANerfstudio

Real-time Pipeline

RGB FrameStandard monocular camera input
TrackingPhotometric pose estimation
MappingKeyframe selection & Gaussian optimization
RenderingNovel view synthesis
No Depth Sensor

Works with standard RGB video only — no LiDAR, stereo, or depth cameras needed

Real-time SLAM

Simultaneous tracking and mapping at interactive frame rates

3D Gaussian Splatting

Photo-realistic rendering via differentiable Gaussian primitives, outperforming NeRF in speed

Nerfstudio Extension

Built as a modular plugin for the Nerfstudio framework, easily extensible

Demo Reconstructions

Room Scene

Dense 3D reconstruction of an indoor room

Kitchen Scene

Complex environment with fine details

Living Room

Large-scale open environment

Bachelor's Thesis

This project formed the foundation of my thesis at Sapienza University of Rome

Motivation

Traditional SLAM systems produce sparse point clouds or geometric meshes. These are useful for localization, but visually far from the actual scene. I wanted to build a system that produces photo-realistic 3D reconstructions from nothing more than a standard RGB camera, in real-time.

3D Gaussian Splatting had just emerged as a breakthrough in neural rendering, achieving photorealistic quality with fast rendering times. But it was designed for offline reconstruction from pre-captured images. I adapted it into a real-time SLAM pipeline.

The Approach

SplatSLAM runs a continuous loop over incoming RGB frames:

Tracking: for each new frame, the system estimates the camera's pose by minimizing photometric error between the observed frame and a rendered view from the current 3D map. No feature detection, no depth sensor, just pixel-level alignment.

Mapping: keyframes are selected using a covisibility heuristic and used to optimize the 3D Gaussian Splatting representation. Each Gaussian stores position, covariance, color (spherical harmonics), and opacity. The optimization incrementally refines these parameters as new views arrive.

Rendering: the final map can be rendered from any viewpoint, producing images that are qualitatively close to the original video: not blocky meshes or scattered points, but continuous, detailed surfaces.

The system is built as an extension to the Nerfstudio framework, adapting the splatfacto offline method into a real-time incremental pipeline.

Results

The system achieves real-time dense reconstruction from monocular RGB video, with reconstruction quality significantly exceeding traditional mesh-based or point cloud-based SLAM methods. The output is a 3D scene that can be freely explored with photo-realistic rendering.

Demonstrations on room, kitchen, and living room sequences show the system handling diverse indoor environments with varying levels of complexity and scale.

Context

This project was developed as part of the excellence program at Sapienza University of Rome and became the foundation for my bachelor's thesis. It directly informed my subsequent work on GPU kernel optimization. The performance bottlenecks I encountered here led me to explore Triton and CUDA programming for dense mapping acceleration.