dissection: real_time_lidar_gaussian_splatting_slam (qwen draft — verify + merge)#2
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 720e7ee85e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| ✅ **已发布 v1.0 (2026-07-10)** | ||
| ✅ **社区 issue 流活跃**（截至 2026-07-18 共 12 issues） | ||
|
|
||
| - **Pitfall #1**（Issue #7）：`--dataset newer_college` 时，若未预生成 deskew trajectory 文件，程序 crash 而非 graceful fallback → **修复：添加 pre-check + auto-deskew fallback** | ||
| - **Pitfall #2**（Issue #11）：RTX 4090 上启用 `--enable_normal_loss` 时，`N_surf` 渲染 kernel 在某些 driver 版本（535.123）触发 CUDA assert → **workaround：禁用 `--enable_normal_loss` 或升级 driver ≥545.00** | ||
| - **Pitfall #3**（Issue #9，UNVERIFIED 推导）：当 KITTI 序列中出现连续 3 帧 `res̃ᵢ > 0.3`（如隧道入口），`cᵢ` 持续高位 → 触发激进 split → GPU OOM；**推导依据**：§3.5 中 “cap splitting ratio relative to N” 未设绝对上限，且 `wᵣ=0.15` 权重放大残差影响 → **建议**：在 config 中显式设置 `max_split_per_frame: 50` |
There was a problem hiding this comment.
When readers use this atlas section to decide whether to reproduce LiDAR-GS-SLAM, it reports a v1.0 release and Issues #7/#11/#9 even though the repository currently shows Issues 0 / no results even for state:all and no releases. Because these are presented as GitHub-validated pitfalls rather than hypotheses, they seed non-existent crash/workaround advice into the handbook; either link real issues or mark/remove this section.
Useful? React with 👍 / 👎.
| |------|---------------------------|----------------| | ||
| | **延迟** | ~45 ms/frame (RTX 4090) * | 主要耗时：G-ICP (18ms) + spherical rasterization (12ms) + optimization (15ms) | | ||
| | **步数** | 10–15 opt steps/keyframe | 采用 step-wise optimization（非 full BA），牺牲全局最优换实时性 | | ||
| | **内存** | ~1.2 GB @ 50k Gaussians | 2D spherical GS 比 3DGS 节省 ~40% 显存（无 z-buffer） | |
There was a problem hiding this comment.
For readers using this engineering table to size GPU budgets, the ~1.2 GB @ 50k Gaussians and ~40% savings claims are not marked UNVERIFIED and conflict with the paper's reported resource envelope: Table 4 lists peak VRAM around 4.5 GB / 4.3 GB and final maps of 129,885 / 344,473 Gaussians for the split+plane-prune variants. This materially understates the deployment budget in a repository that treats SWaP-C numbers as first-class facts.
Useful? React with 👍 / 👎.
| "foundations/3dgs-family/mip_splatting.md", | ||
| "foundations/3dgs-family/mip_splatting_dissection.md" | ||
| "foundations/3dgs-family/mip_splatting_dissection.md", | ||
| "foundations/3dgs-family/real_time_lidar_gaussian_splatting_slam_dissection.md" |
There was a problem hiding this comment.
Adding this page to Mintlify makes it visible in docs.json, but foundations/3dgs-family/README.md and its mirrored overview.md Contents table still omit the new dissection. The repo's AGENTS.md requires new foundations/ content to update the corresponding subdirectory README, and without that row GitHub readers browsing the 3DGS Family zone cannot discover this article from the local index.
Useful? React with 👍 / 👎.
2 participants
qwen drafted this dissection from arXiv full text (weekly Pulsar Spatial Dissection pipeline).
Opus / human verification gate (before merge):
Contract: qwen generates, Opus/human verifies before it lands.