Tips for large FOV fisheye lens

[LaFeuilleMorte]

Hi, I'm using your fisheye implementation to reconstruct from fisheye images. The results were quite good with smaller FOV fisheye lens, Here's the results from zipnerf berlin datasets:

But the results got worse when I'm using my own datasets which has a much larger fov, the results become really bad.

Here's an image of my dataset:

And here's the rendering results on eval set:

Due to there'are some ring artifacts on the edge of the fisheye images. I simply apply a mask on the inupt images and do colmap and GS reconstruction on the masked images. My mask was :

[scott198510]

@LaFeuilleMorte What you said is indeed correct. The fisheye-gs model (gsplat refer to ) is based on equiaxed projection to consider distortion processing, and the resulting effect has a relatively large projection error, especially for images with high distortion and corner areas. You can refer to the following three papers, which are more friendly to processing large distortion fisheye images
1st: https://arxiv.org/html/2402.00752v3
2nd: https://arxiv.org/abs/2402.00763
3rd: https://research.nvidia.com/labs/toronto-ai/3DGUT

Anyway, I have also been paying attention to using high distortion fisheye images for 3DGS。

[scott198510]

Hi, I'm using your fisheye implementation to reconstruct from fisheye images. The results were quite good with smaller FOV fisheye lens, Here's the results from zipnerf berlin datasets:

@LaFeuilleMorte

[LaFeuilleMorte]

https://arxiv.org/html/2402.00752v3

Thank you sincerely for your advice. I'll read these papers!