Exploring the Efficiency of Gaussian Splats for 3D scene Representation
Representing 3D scenes efficiently is a constant challenge in computer graphics. Recently, Gaussian splats have emerged as a promising technique, offering a compelling balance between rendering quality and performance. I’ve been exploring their behaviour, particularly when applied to distinctly different types of scenes – and the results are quite insightful.
Here’s a look at how Gaussian splats perform with varying levels of detail and complexity, using a “caterpillar” model as a case study.The following table summarizes the key metrics for different resolutions of this model:
| Resolution | Rendering Time | Number of Splats | File Size (Uncompressed) | File Size (Compressed) |
|———————-|—————-|——————|————————–|————————-|
| 320×180 | 2m 39s | 1,105,383 | 274 MB | 24.5 MB |
| 920×512 | 6m 48s | 1,611,563 | 400 MB | 36.4 MB |
| 1940×1080 | 15m 27s | 2,046,676 | 507 MB | 46 MB |
| Original (1957×1090) | 15m 9s | 900,798 | 223 MB | 21 MB |
These numbers reveal some interesting trends. Generally, as resolution increases, both rendering time and the number of splats required also increase. However, the relationship isn’t perfectly linear.
Notably, a line drawing scene requires roughly double the number of splats and file size compared to its source scene. I hypothesize this occurs because Gaussian splats excel at modeling large areas and textures. Consequently, rendering thin lines demands a greater number of individual Gaussians to accurately represent those features.
Essentially, splats are more efficient at representing continuous surfaces than discrete lines. This suggests that different representation strategies might be optimal depending on the scene’s characteristics. You might consider hybrid approaches that leverage the strengths of Gaussian splats alongside other techniques for representing linear features.
If you’re interested in experimenting with this yourself, the code to generate these results is a combination of scripts designed to work with the referenced libraries. Feel free to reach out if you’d like to discuss running these experiments or exploring potential collaborations.
I’m always eager to hear your thoughts on this work and explore new avenues for research. You can connect with me at tansh at amritkwatra dot com.
I’d like to acknowledge the valuable feedback and suggestions from Ritik Batra,ilan mandel,and Thijs Roumen,which greatly contributed to this exploration.