3D Occupancy Perception: a curated survey of dense scene understanding

A survey that is also a live paper list

The repository is described in its own summary as research into 3D occupancy perception for autonomous driving, covering LiDAR-centric, vision-centric and multi-modal occupancy perception with information fusion techniques discussed, and it states the aim of being the most comprehensive survey to date on the topic. The README makes clear that this is an active repository rather than a frozen artefact, that readers are invited to watch it for the latest advances, and that the authors want contributions of related work, including applications and not only perception, with pull requests and a contact route offered. That is the key structural decision. A survey paper is a snapshot, but the repository that supports it is a living index, and the news log shows the difference, with entries recording the acceptance by Information Fusion on 2024-09-03, more representative works and benchmarking comparisons incorporated on 2024-07-21 bringing the total to 192 literature references, more figures added and occupancy-based applications reorganised on 2024-05-18, and the first version of the survey available on arXiv on 2024-05-08.

The framing that separates occupancy from BEV

The Introduction states the core argument, and it is the reason the survey exists as a separate body of work. 3D occupancy perception aims to observe and understand dense 3D environments for autonomous vehicles, and because of its comprehensive perception capability it is emerging as a trend in autonomous driving perception systems, attracting attention from industry and academia. The technical distinction the paper draws is against traditional bird's-eye view perception. Occupancy perception has the same nature of multi-source input and the same necessity for information fusion as BEV, but the difference is that it captures vertical structures that 2D BEV ignores. That single point is what the taxonomy, the fusion techniques and the training discussion are all downstream of, and it explains why a survey devoted to information fusion for this problem is framed from the fusion perspective in the first place.

The taxonomy is by sensor, and the survey leads with LiDAR

The table of contents splits the methods section into LiDAR-centric, vision-centric, radar-centric and multi-modal occupancy perception, followed by sections on 3D occupancy datasets and on occupancy-based applications, then the citation and contact sections. The LiDAR table shown in the README is the format every modality table follows, with a year, a venue, a paper title, a link and a code or project page column, and it is worth reading as a signal of where the field is moving. The 2026 entries include Semantic Occupancy Prediction with Dual Range-Voxel Representation at arXiv with a code link, TFusionOcc, a student's t-distribution based object-centric multi-sensor fusion framework, and LiFlow, flow matching for 3D LiDAR scene completion. The 2025 entries cover octree latent diffusion for semantic 3D scene generation and completion, and diffusion distillation with direct preference optimization for efficient 3D LiDAR scene completion. The 2024 entries include TALoS at NeurIPS, PaSCo at CVPR, LiDAR-based 4D occupancy completion and forecasting at IROS, and OccRWKV, which targets linear complexity for efficient 3D semantic occupancy prediction.

What the survey claims to cover, in its own words

The highlight section lists three claims and they are specific enough to check. First, that this is a systematic survey of the latest research on 3D occupancy perception in the field of autonomous driving. Second, that the survey provides the taxonomy of 3D occupancy perception and elaborates on core methodological issues, naming network pipelines, multi-source information fusion and effective network training as the three. Third, that the survey presents evaluations for 3D occupancy perception with detailed performance comparisons, and that current limitations and future research directions are discussed. That third claim is the one that matters most for a practitioner, because a taxonomy tells you how the field is organised while an evaluation table tells you which of them actually works on a given dataset. The Introduction says the same thing in the methods, that it reviews recent works with in-depth analyses of methodologies across input modalities, summarises general network pipelines, highlights information fusion techniques and discusses effective network training, then evaluates the occupancy perception performance of the state of the art on the most popular datasets.

How to use the repository, and what it is not

The practical entry point is the README itself, which links the survey PDF on arXiv, the author's scholarly profiles and the project homepage. The repository's top level is minimal, a README and an `assets/` directory holding the figures, including the title graphic, a diagram of an autonomous driving vehicle system and a brief history figure, and the tables carry links out to arXiv, to code repositories and to project pages. So the usage pattern is: read the survey for the argument, use the tables to find the paper for your modality, follow the code link if you want to reproduce it. The limits are equally clear. There is no implementation here, no training script, no dataset loader, and the primary language of the repository is recorded as unknown, which for a repository whose only tracked file is Markdown and images is an artefact of metadata rather than a signal. The evaluation data is also second-hand, since the performance comparisons in the survey are the results other groups reported, and the repository itself offers no benchmark harness of its own.

Where this kind of curated list falls short

Three limits are worth naming. The first is the tension between a live list and a published paper, and it has a practical consequence. The news log records the first arXiv version on 2024-05-08 and the acceptance on 2024-09-03, while the tables in the README carry 2026 entries, so the curated list is well ahead of the published survey, and a reader who wants the peer reviewed argument has to work with a version that predates most of the table. The second is that a curated list reflects its curators' judgement about what to include, and the invite to submit missing work is an open door, which means coverage is a function of who has time to send pull requests rather than a systematic search protocol. The third is the modality balance. The LiDAR table is populated with code links and project pages for most entries, and there is no reason to expect the same density for vision, radar or multi-modal work, so the apparent richness of the field in one modality and the apparent thinness in another may be partly a reflection of curation effort rather than of research volume. None of this makes the repository less useful; it makes it a map with a date on it.

Editorial conclusion

The repository is worth bookmarking for anyone entering 3D occupancy perception, because it is a maintained, link-rich index with code attached to most entries rather than a bibliography, and because the survey it tracks was accepted by a journal with an impact factor the README puts at 14.7. It is the wrong tool if you want an implementation, since the repository holds a README and an assets directory and no code, and the wrong tool if you want a definitive ranking, since the survey's own comparison tables are the only evaluation data offered. Before you rely on it, check the survey's date against the current state of the field, since the curated list keeps growing while the published version does not, and read the taxonomy section rather than jumping to a modality chapter. The last push was on 2026-09-02 and the repository publishes no GitHub releases.

Frequently asked questions

Official sources

1. HuaiyuanXu/3D-Occupancy-Perception on GitHub
2. Issues
3. Project website
4. README