[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80586":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":8,"htmlUrl":8,"language":9,"languages":8,"totalLinesOfCode":8,"stars":10,"forks":11,"watchers":12,"openIssues":13,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":15,"stars7d":16,"stars30d":17,"stars90d":14,"forks30d":14,"starsTrendScore":18,"compositeScore":19,"rankGlobal":8,"rankLanguage":8,"license":20,"archived":21,"fork":21,"defaultBranch":22,"hasWiki":23,"hasPages":21,"topics":24,"createdAt":8,"pushedAt":8,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":14,"starSnapshotCount":14,"syncStatus":13,"lastSyncTime":28,"discoverSource":29},80586,"R3","KevinXu02\u002FR3","KevinXu02",null,"Python",187,14,4,2,0,12,36,99,47,3.53,"Apache License 2.0",false,"main",true,[],"2026-06-12 02:04:04","\u003Cdiv align=\"center\">\n\u003Ch1 style=\"border-bottom: none; margin-bottom: 0px \">R³: 3D Reconstruction via Relative Regression\u003C\u002Fh1>\n\n\u003Ca href='https:\u002F\u002Fkevinxu02.github.io\u002Fr3-site\u002F'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject_Page-R3-green' alt='Project Page'>\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Fhuggingface.co\u002FKevinXu02\u002FR3'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97%20Hugging%20Face-Checkpoints-blue' alt='Checkpoints'>\u003C\u002Fa>\n\u003Ca href='https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.26519'>\u003Cimg src='https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FarXiv-2605.26519-b31b1b' alt='arXiv'>\u003C\u002Fa>\n\n\u003C\u002Fdiv>\n\nThis work presents **R³**, a feed-forward model that reconstructs camera poses and dense geometry from arbitrarily long video streams via *relative-pose regression*. Instead of regressing every camera in one global frame, R³ predicts confidence-weighted pairwise relative poses on top of a Depth Anything 3 backbone, then assembles a consistent global trajectory downstream.\n\nTwo ideas keep the modeling minimal:\n\n- A **lightweight pairwise pose MLP** sits on a DA3 backbone — no recurrent state, no TTT modules, no extra transformer.\n- A **single learned confidence per edge** (decoupled into rotation and translation) drives loss weighting, pose aggregation, and keyframe-bank management.\n\nWith **372M parameters** (≈⅓ of recent 1B-class models), R³ matches or surpasses state-of-the-art streaming methods on pose estimation and dense reconstruction, runs at **20+ FPS**, and scales to **thousands of frames** under a bounded memory budget.\n\n## 📰 News\n\n- **2026-05-26:** Inference-only public release with `r3` and `r3_long` checkpoints.\n\n## 🗂️ Release TODO\n\n- [ ] Evaluation code.\n- [ ] Training code.\n\n## 🚀 Quick Start\n\n### 📦 Installation\n\n```bash\nconda env create -f environment.yml\nconda activate r3\npip install -e .\n```\n\nIf you already have a CUDA-enabled PyTorch environment, install dependencies directly:\n\n```bash\npip install -r requirements.txt\npip install -e .\n```\n\n### 🧱 Checkpoints\n\nPlace weights under:\n\n```text\nckpt\u002Fr3.safetensors\nckpt\u002Fr3_long.safetensors\n```\n\nBoth are available on [Hugging Face](https:\u002F\u002Fhuggingface.co\u002FKevinXu02\u002FR3):\n\n| Name      | File                                                                                       | Train views | Best for                       | Notes                                                                                                       |\n|-----------|--------------------------------------------------------------------------------------------|-------------|--------------------------------|-------------------------------------------------------------------------------------------------------------|\n| `r3`      | [`r3.safetensors`](https:\u002F\u002Fhuggingface.co\u002FKevinXu02\u002FR3\u002Fblob\u002Fmain\u002Fr3.safetensors)           | 4–32        | Indoor \u002F small-coverage scenes | Default checkpoint, reported in the paper. Stronger local consistency on short clips.                       |\n| `r3_long` | [`r3_long.safetensors`](https:\u002F\u002Fhuggingface.co\u002FKevinXu02\u002FR3\u002Fblob\u002Fmain\u002Fr3_long.safetensors) | 32–100      | Outdoor \u002F long trajectories    | Used by `--mode long` and `--mode strided` unless `--ckpt` is passed explicitly.                             |\n\n### 💻 Run the Demo\n\n```bash\npython demo.py --seq_path examples\u002Findoor --no_viewer\n```\n\n`demo.py` runs inference with `infer.py`, writes depth \u002F color \u002F confidence \u002F camera files to `--output_dir`, and then opens the saved run in a [Viser](https:\u002F\u002Fviser.studio\u002F) viewer.\nBy default, the demo uses `--mode test`, which keeps all KV entries and skips fallback \u002F metric scale for a quick smoke run. Use `--mode local`, `--mode long`, or `--mode strided` for the release presets.\nSky-mask export is available through `--compute_sky_mask` when the selected model emits a `sky` tensor; the default R3 checkpoint does not emit one.\n\nPresets cover the common regimes:\n\n```bash\npython demo.py --mode test     # quick test run, all KV cache\npython demo.py --mode local    # indoor scenes, small coverage\npython demo.py --mode long     # long trajectories, large outdoor scenes\npython demo.py --mode strided  # temporally strided video\n```\n\n(`--mode short`, `--mode sampled`, and `--mode sparse` are kept as legacy aliases.)\n\nTo reopen a saved run without re-running inference:\n\n```bash\npython view.py --data_dir scratch\u002Fdemo\u002F\u003Crun_name>\n```\n\n## 🙏 Acknowledgement\n\nOur code is built upon the following repositories:\n\n- [Depth Anything 3](https:\u002F\u002Fgithub.com\u002FByteDance-Seed\u002FDepth-Anything-3)\n- [CUT3R](https:\u002F\u002Fgithub.com\u002FCUT3R\u002FCUT3R)\n- [STream3R](https:\u002F\u002Fgithub.com\u002FNIRVANALAN\u002FSTream3R)\n\nWe thank the authors for their excellent work.\n\n## 📝 Citation\n\nIf R³ is useful in your research or projects, please cite:\n\n```bibtex\n@article{r3_2026,\n  title  = {R^3: 3D Reconstruction via Relative Regression},\n  author = {Anonymous},\n  year   = {2026},\n  note   = {Paper coming soon}\n}\n```\n\nPlease also cite the works above if you use this codebase.\n","R³是一个基于相对回归的3D重建模型，能够从任意长度的视频流中重构相机姿态和密集几何。其核心功能包括通过轻量级的成对姿态MLP预测置信度加权的相对姿态，并在Depth Anything 3骨干网络上进行全局轨迹组装。该模型仅使用372M参数就能匹配或超越现有的流式方法，在保持高精度的同时以超过20 FPS的速度运行，并且可以处理数千帧的数据，适用于室内或小范围场景的3D重建任务。R³的设计简洁高效，没有复杂的循环状态或额外的变换器模块，适合需要实时或近实时3D重建的应用场景。","2026-06-11 04:01:17","CREATED_QUERY"]