heshuting555/Awesome-3DGS-Applications

What it solves

3D Gaussian Splatting (3DGS) is a technique originally built for view synthesis — rendering new camera angles of a captured scene. But researchers have since pushed it into other directions. This project addresses the lack of a single organized reference for those newer uses. It is the first dedicated survey covering 3DGS applications in three rapidly evolving areas: segmentation (picking out objects or regions in a 3D scene), editing (modifying a reconstructed 3D scene), and generation (creating new 3D content from images or text). The survey paper has been accepted to TPAMI.

How it works

The repository is a structured, table-based catalog. It lists:

• Datasets for segmentation (e.g. ScanNet, Replica, LERF variants, SceneSplat), each with venue, scene count, view count, and a one-line highlight of what makes it useful.
• Datasets for editing (e.g. DTU, Tanks and Temples, LLFF, ScanNet++), again with the same metadata.
• Datasets for generation (e.g. ShapeNet, Objaverse, LAION-5B, DL3DV-10K), including the type of task they support (image-to-3D or text-to-3D).
• Methods for segmentation — a long list of papers (2023–2025) with venue, paper title, and a link to either the paper or code when available.
• Methods for editing and methods for generation, also cited.
• Other application tasks: human avatar modeling, SLAM, and 3DGS-based detection.
• A pointer to related surveys and a citation entry.

The tables make it easy to scan what exists, where it was published, and whether source code is available.

Who it's for

Researchers, graduate students, and practitioners working in 3D computer vision, graphics, or scene understanding who want a quick map of the 3DGS application landscape. It is also a quick reference for choosing datasets or baseline methods when starting new experiments. The maintainers explicitly invite the community to suggest missing work via issues or pull requests.

Highlights

• First dedicated survey on 3DGS downstream applications beyond classical view synthesis
• Accepted to TPAMI (IEEE Transactions on Pattern Analysis and Machine Intelligence)
• Covers three main directions — segmentation, editing, and generation — in depth
• Includes detailed dataset tables with scene counts, view counts, and purpose notes
• Lists dozens of methods with direct links to papers and code repositories
• Covers extra application areas: human avatars, SLAM, and 3DGS-based detection
• Open to community contributions for missing or new work