Efficient Rendering of Dynamic Scenes: A New Approach with Gaussian Splatting

Rendering dynamic scenes from monocular videos is a central challenge in computer graphics and computer vision. Recent advancements in Deformable Gaussian Splatting have yielded promising results by robustly representing real-world dynamic scenes. However, this method often leads to a high redundancy of Gaussian functions as it attempts to capture each training view at different time points. This results in slower rendering speeds. Furthermore, the attributes of Gaussian functions in static regions are time-invariant, making the modeling of each individual Gaussian function unnecessary and potentially leading to unwanted flickering in these areas. The number of Gaussian functions represents the biggest bottleneck for the rendering speed of dynamic scenes in practice.

A new approach, known as Efficient Dynamic Gaussian Splatting (EDGS), promises a remedy. EDGS represents dynamic scenes through sparse, time-varying attribute modeling. Instead of modeling each Gaussian function individually, EDGS uses a sparse anchor grid representation. The motion flow of the dense Gaussian functions is calculated via a classic kernel representation. This approach significantly reduces the number of Gaussian functions to be processed, thus accelerating the rendering process.

Another important aspect of EDGS is the efficient filtering of anchors corresponding to static regions. Through an unsupervised strategy, anchors related to deformable objects are identified, and only these are fed into MLPs (Multi-Layer Perceptrons) to query time-varying attributes. Anchors in static regions, on the other hand, are ignored, further reducing the computational load and minimizing flickering in these areas.

Experiments with real-world datasets show that EDGS significantly improves rendering speed compared to previous state-of-the-art methods while achieving superior rendering quality. Through its sparse representation and targeted modeling of time-varying attributes, EDGS offers an efficient and effective way to represent dynamic scenes.

The Advantages of EDGS at a Glance:

EDGS addresses the challenges of conventional Gaussian Splatting by:

Reducing the number of Gaussian functions through a sparse anchor grid representation.

Efficiently modeling the motion flow through a classic kernel representation.

Unsupervised filtering of anchors in static regions to minimize computational load and flickering.

Targeted querying of time-varying attributes only for anchors associated with deformable objects.

The combination of these techniques allows for a significant improvement in rendering speed and quality when representing dynamic scenes.

Outlook and Future Developments

Research in the field of dynamic scene rendering is dynamic and promising. EDGS represents an important step towards more efficient and higher-quality rendering of dynamic scenes. Future work could focus on further optimizing anchor placement and developing even more robust methods for identifying dynamic regions. Integrating EDGS into existing rendering pipelines and applying it to various scenarios, such as augmented reality and virtual environments, also offers great potential.