Only the front data preprocessing part is included in the code, the 3DGS part is not included and you need to build it yourself. That is, only the following parts are included.
2. Comparative analysis with street gaussian
cap (a poem)Street Gaussians: Modeling Dynamic Urban Scenes with Gaussian Splatting, also from Li auto's work, is the work of zju and Li auto together, the two are actually quite different.
Common Ground:
Both modeled objects in 3d gaussian; both processed adas scenes; both used additional point cloud instead of only SFM generated.
Points of difference:
street gaussian is modeled in the scene, is subtracted from the background leaving the foreground, the background is modeled in one wayIntroduction to Spherical Harmonic Functions（Spherical Harmonics）Prospects are modeled in one waydynamic spherical
harmonics model dynamic (science)Introduction to Spherical Harmonic Functions（Spherical Harmonics）The process is therefore promising and contextualized. Thus, there is a foreground and a background to the process. The background was modeled and the foreground was modeled. The results are rather vague, as if there is no way to take out a single target? This way is scene reconstruction, but you can't take out a single target and put it to use somewhere else, or you don't know how to use it?
3drealdata is a form of segmentation that removes the background and leaves the foreground, which is then modeled using 3d gaussian. Only the foreground target is reconstructed. The same target image is richer, the result is clearer, and can be rotated to a large angle, and the data generated in this way can be used by putting the vehicle in any position or at any pose. 3d real data uses the standard gaussian splatting, while street gaussian is improved.