[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80617":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":12,"lastSyncTime":28,"discoverSource":29},80617,"Open-d4rt","Lijiaxin0111\u002FOpen-d4rt","Lijiaxin0111",null,"Python",389,14,2,6,0,267,288,309,801,3.53,"Apache License 2.0",false,"main",true,[],"2026-06-12 02:04:04","\u003Cdiv align=\"center\">\n  \u003Ch1>OpenD4RT\u003C\u002Fh1>\n  \u003Ch3>An unofficial PyTorch\u002FGPU implementation of D4RT for 4D reconstruction and tracking\u003C\u002Fh3>\n  \u003Cp>\n    \u003Cstrong>RHOS Team\u003C\u002Fstrong> · \u003Ca href=\"https:\u002F\u002Fmvig-rhos.com\u002F\" target=\"_blank\">https:\u002F\u002Fmvig-rhos.com\u002F\u003C\u002Fa>\n  \u003C\u002Fp>\n  \u003Cp>\n    \u003Ca href=\"https:\u002F\u002Fd4rt-paper.github.io\u002F\" target=\"_blank\">\n      \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%8C%90-D4RT%20Project-2f80ed\" alt=\"D4RT project page\">\n    \u003C\u002Fa>\n    \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002FLijiaxin0111\u002FOpenD4RT\u002Ftree\u002Fmain\u002Fcheckpoints\" target=\"_blank\">\n      \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002F%F0%9F%A4%97-Checkpoints-yellow\" alt=\"Hugging Face checkpoints\">\n    \u003C\u002Fa>\n    \u003Ca href=\"LICENSE\">\n      \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Flicense-Apache%202.0-blue.svg\" alt=\"License\">\n    \u003C\u002Fa>\n    \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpython-3.10-blue.svg\" alt=\"Python\">\n    \u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPyTorch-2.6-red.svg\" alt=\"PyTorch\">\n  \u003C\u002Fp>\n  \u003Cp>\u003Cstrong>OpenD4RT reproduces D4RT-style 4D reconstruction and tracking with released WorldTrack evaluation, visualization tools, and Hugging Face checkpoints.\u003C\u002Fstrong>\u003C\u002Fp>\n\u003C\u002Fdiv>\n\nOpenD4RT is an unofficial open-source PyTorch\u002FGPU implementation of D4RT,\ndeveloped to reproduce the model architecture, training recipe, evaluation\nprotocols, and implementation details described in the D4RT paper and\nappendix. The current public repo includes the released Hugging Face\ncheckpoint, the model, WorldTrack evaluation, and Viser visualization\ntools, with complete training and evaluation code planned for release.\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"docs\u002Fimage.png\" width=\"950\" alt=\"D4RT overview\">\n\u003C\u002Fp>\n\n## 🔥 News\n\n- [2026\u002F05\u002F02] Released the OpenD4RT WorldTrack evaluation pipeline, Viser\n  visualization tools, and the first Hugging Face checkpoint.\n\n## 🧠 What is D4RT?\n\nD4RT is a feedforward video model for reconstructing and tracking dynamic\nscenes. It uses a unified transformer architecture to infer depth,\nspatio-temporal correspondence, and camera parameters from a single video. Its\nquery interface probes the 3D position of a source pixel `(u, v, t_src)` at a\ntarget timestep `t_tgt` in a selected camera coordinate frame `t_cam`, enabling\nsparse tracking, all-pixel tracking, and 4D scene reconstruction through the\nsame model interface.\n\nSee [docs\u002FD4RT_paper.pdf](docs\u002FD4RT_paper.pdf) for the local paper PDF\nincluded in this repository.\n\n## 🔧 Installation\n\nCreate the conda environment:\n\n```bash\nconda env create -f environment.yml\nconda activate d4rt\n```\n\nOr install into an existing Python environment:\n\n```bash\npip install -r requirements.txt\n```\n\nThe visualization package builder calls the `ffmpeg` command-line tool to\nwrite MP4 assets for Viser. The conda environment includes `ffmpeg`; if you use\n`pip install -r requirements.txt`, install `ffmpeg` separately if needed.\n\n## 📦 Checkpoint Zoo\n\n| Variant | Data | Aug. | Frames | Status | Download |\n| --- | --- | --- | ---: | --- | --- |\n| `32CLIP_9Dataset_NoAUG` | 9Mix |  color aug + No crop aug | 32 | Released | [HF](https:\u002F\u002Fhuggingface.co\u002FLijiaxin0111\u002FOpenD4RT\u002Ftree\u002Fmain\u002Fcheckpoints\u002FOpenD4RT_32CLIP_9Dataset_NoAUG) |\n| `48CLIP_9Mix_NoCropAUG` | 9Mix | color aug + No crop aug  | 48 | Released | [HF](https:\u002F\u002Fhuggingface.co\u002FLijiaxin0111\u002FOpenD4RT\u002Ftree\u002Fmain\u002Fcheckpoints\u002FOpenD4RT_48CLIP_9Mix_NoCropAUG) |\n| `48CLIP_9Mix_AUG` | 9Mix | color aug + crop aug | 48 | Coming | TBD |\n| `32CLIP_10Mix_SynthVerse_NoAUG` | 10Mix | color aug + No crop aug | 32 | Coming | TBD |\n| `48CLIP_10Mix_SynthVerse_AUG` | 10Mix |  color aug + crop aug | 48 | Coming | TBD |\n\nReleased checkpoint local path:\n`checkpoints\u002FOpenD4RT_32CLIP_9Dataset_NoAUG\u002Fopend4rt.ckpt`.\n\nAdditional released checkpoint local path:\n`checkpoints\u002FOpenD4RT_48CLIP_9Mix_NoCropAUG\u002Fopend4rt.ckpt`.\n\nTip: all rows are OpenD4RT variants. The 9Mix setting uses PointOdyssey,\nDynamic Replica, Kubric Full,\nTartanAir, Virtual KITTI 2, ScanNet, BlendedMVS, CO3D, and MVS-Synth. The\n10Mix setting additionally includes SynthVerse.\n\n## ⬇️ Checkpoint Download\n\nDownload the released checkpoint and model config from\n[Lijiaxin0111\u002FOpenD4RT](https:\u002F\u002Fhuggingface.co\u002FLijiaxin0111\u002FOpenD4RT\u002Ftree\u002Fmain\u002Fcheckpoints)\ninto the default path used by the scripts:\n\n```bash\npip install -U huggingface_hub\n\nhuggingface-cli download Lijiaxin0111\u002FOpenD4RT \\\n  --repo-type model \\\n  --include \"checkpoints\u002FOpenD4RT_32CLIP_9Dataset_NoAUG\u002Fopend4rt.ckpt\" \\\n  --include \"checkpoints\u002FOpenD4RT_32CLIP_9Dataset_NoAUG\u002Fmodel.yaml\" \\\n  --include \"checkpoints\u002FOpenD4RT_48CLIP_9Mix_NoCropAUG\u002Fopend4rt.ckpt\" \\\n  --include \"checkpoints\u002FOpenD4RT_48CLIP_9Mix_NoCropAUG\u002Fmodel.yaml\" \\\n  --local-dir .\n```\n\nExpected local files:\n\n```text\ncheckpoints\u002FOpenD4RT_32CLIP_9Dataset_NoAUG\u002F\n  opend4rt.ckpt\n  model.yaml\ncheckpoints\u002FOpenD4RT_48CLIP_9Mix_NoCropAUG\u002F\n  opend4rt.ckpt\n  model.yaml\n```\n\n## 🌍 WorldTrack Data\n\nDownload the WorldTrack release from:\n\n```text\nhttps:\u002F\u002Fdrive.google.com\u002Fdrive\u002Ffolders\u002F1-JW88ru30irMYyFab_4YBQbGbd9tKpXV\n```\n\nPlace the `.npz` files under:\n\n```text\ndata\u002Fworldtrack_release\u002F\n  adt_mini\u002F*.npz\n  po_mini\u002F*.npz\n  pstudio_mini\u002F*.npz\n  ds_mini\u002F*.npz\n```\n\n## 📊 Evaluation\n\nRun a quick smoke test on one `adt_mini` sequence:\n\n```bash\nLIMIT_SEQS=1 SUBSETS=adt_mini OUTPUT_DIR=tmp\u002Feval_smoke bash run_eval_worldtrack.sh\n```\n\nRun the full WorldTrack evaluation:\n\n```bash\nbash run_eval_worldtrack.sh\n```\n\nEquivalent explicit command:\n\n```bash\nEXP=checkpoints\u002FOpenD4RT_32CLIP_9Dataset_NoAUG\n\npython eval_track3d_in_worldtrack.py \\\n  --model-config \"$EXP\u002Fmodel.yaml\" \\\n  --ckpt-path \"$EXP\u002Fopend4rt.ckpt\" \\\n  --data-root data\u002Fworldtrack_release \\\n  --subsets adt_mini,po_mini,pstudio_mini,ds_mini \\\n  --num-frames 64 \\\n  --query-chunk-size 4096 \\\n  --output-dir tmp\u002Feval_worldtrack \\\n  --device cuda \\\n  --save-per-sequence\n```\n\nUseful overrides:\n\n```bash\nQUERY_CHUNK_SIZE=1024 bash run_eval_worldtrack.sh\nCUDA_VISIBLE_DEVICES=1 DEVICE=cuda bash run_eval_worldtrack.sh\nSUBSETS=adt_mini LIMIT_SEQS=1 NUM_FRAMES=64 bash run_eval_worldtrack.sh\n```\n\n## 🏆 Results\n\nOpenD4RT_32CLIP_9Dataset_NoAUG detailed WorldTrack results:\n\n| Subset | APD global | EPE global | APD global dyn | EPE global dyn | Queries |\n| --- | ---: | ---: | ---: | ---: | ---: |\n| `adt_mini` | 0.6993 | 0.2964 | 0.6975 | 0.3628 | 22187 |\n| `po_mini` | 0.6603 | 0.3397 | 0.7333 | 0.2722 | 53468 |\n| `pstudio_mini` | 0.7863 | 0.1811 | 0.7863 | 0.1811 | 8720 |\n| `ds_mini` | 0.7266 | 0.2944 | 0.7521 | 0.2699 | 52462 |\n\nOpenD4RT_48CLIP_9Mix_NoCropAUG detailed WorldTrack results\n(`step_0006000`, `anchor_clip`, evaluated with 64 frames):\n\n| Subset | APD global | EPE global | APD global dyn | EPE global dyn | Queries |\n| --- | ---: | ---: | ---: | ---: | ---: |\n| `adt_mini` | 0.7220 | 0.2758 | 0.7325 | 0.3199 | 22187 |\n| `po_mini` | 0.6799 | 0.3178 | 0.7425 | 0.2593 | 53468 |\n| `pstudio_mini` | 0.7960 | 0.1753 | 0.7960 | 0.1753 | 8720 |\n| `ds_mini` | 0.7248 | 0.2959 | 0.7488 | 0.2755 | 52462 |\n\n## 📈 Model Results\n\nSparse point tracking comparison on WorldTrack-style subsets. APD is shown as\na percentage, higher APD is better, and lower EPE is better. Recent baseline\nnumbers are transcribed from the sparse point tracking table in the provided\nreference image. OpenD4RT uses this repository's evaluation results, with\n`ds_mini` reported in the DR column.\n\n\u003Ctable>\n  \u003Cthead>\n    \u003Ctr>\n      \u003Cth rowspan=\"2\" align=\"left\">Model\u003C\u002Fth>\n      \u003Cth colspan=\"2\" align=\"center\">PStudio\u003C\u002Fth>\n      \u003Cth colspan=\"2\" align=\"center\">PO\u003C\u002Fth>\n      \u003Cth colspan=\"2\" align=\"center\">DR\u003C\u002Fth>\n      \u003Cth colspan=\"2\" align=\"center\">ADT\u003C\u002Fth>\n    \u003C\u002Ftr>\n    \u003Ctr>\n      \u003Cth align=\"right\">APD ↑\u003C\u002Fth>\u003Cth align=\"right\">EPE ↓\u003C\u002Fth>\n      \u003Cth align=\"right\">APD ↑\u003C\u002Fth>\u003Cth align=\"right\">EPE ↓\u003C\u002Fth>\n      \u003Cth align=\"right\">APD ↑\u003C\u002Fth>\u003Cth align=\"right\">EPE ↓\u003C\u002Fth>\n      \u003Cth align=\"right\">APD ↑\u003C\u002Fth>\u003Cth align=\"right\">EPE ↓\u003C\u002Fth>\n    \u003C\u002Ftr>\n  \u003C\u002Fthead>\n  \u003Ctbody>\n    \u003Ctr>\u003Ctd>\u003Cb>SpaTrackerV2\u003C\u002Fb> (2025)\u003C\u002Ftd>\u003Ctd align=\"right\">74.16\u003C\u002Ftd>\u003Ctd align=\"right\">0.2272\u003C\u002Ftd>\u003Ctd align=\"right\">69.57\u003C\u002Ftd>\u003Ctd align=\"right\">0.3780\u003C\u002Ftd>\u003Ctd align=\"right\">73.43\u003C\u002Ftd>\u003Ctd align=\"right\">0.2732\u003C\u002Ftd>\u003Ctd align=\"right\">92.22\u003C\u002Ftd>\u003Ctd align=\"right\">0.0915\u003C\u002Ftd>\u003C\u002Ftr>\n    \u003Ctr>\u003Ctd>\u003Cb>St4RTrack\u003C\u002Fb> (2025)\u003C\u002Ftd>\u003Ctd align=\"right\">69.67\u003C\u002Ftd>\u003Ctd align=\"right\">0.2637\u003C\u002Ftd>\u003Ctd align=\"right\">67.95\u003C\u002Ftd>\u003Ctd align=\"right\">0.3140\u003C\u002Ftd>\u003Ctd align=\"right\">73.74\u003C\u002Ftd>\u003Ctd align=\"right\">0.2682\u003C\u002Ftd>\u003Ctd align=\"right\">76.01\u003C\u002Ftd>\u003Ctd align=\"right\">0.2680\u003C\u002Ftd>\u003C\u002Ftr>\n    \u003Ctr>\u003Ctd>\u003Cb>TraceAnything\u003C\u002Fb> (2025)\u003C\u002Ftd>\u003Ctd align=\"right\">71.33\u003C\u002Ftd>\u003Ctd align=\"right\">0.2727\u003C\u002Ftd>\u003Ctd align=\"right\">39.83\u003C\u002Ftd>\u003Ctd align=\"right\">1.0593\u003C\u002Ftd>\u003Ctd align=\"right\">60.63\u003C\u002Ftd>\u003Ctd align=\"right\">0.5758\u003C\u002Ftd>\u003Ctd align=\"right\">75.65\u003C\u002Ftd>\u003Ctd align=\"right\">0.2511\u003C\u002Ftd>\u003C\u002Ftr>\n    \u003Ctr>\u003Ctd>\u003Cb>Any4D\u003C\u002Fb> (2025)\u003C\u002Ftd>\u003Ctd align=\"right\">60.03\u003C\u002Ftd>\u003Ctd align=\"right\">0.3344\u003C\u002Ftd>\u003Ctd align=\"right\">60.86\u003C\u002Ftd>\u003Ctd align=\"right\">0.4194\u003C\u002Ftd>\u003Ctd align=\"right\">68.39\u003C\u002Ftd>\u003Ctd align=\"right\">0.3012\u003C\u002Ftd>\u003Ctd align=\"right\">56.71\u003C\u002Ftd>\u003Ctd align=\"right\">0.4320\u003C\u002Ftd>\u003C\u002Ftr>\n    \u003Ctr>\u003Ctd>\u003Cb>V-DPM\u003C\u002Fb> (2026)\u003C\u002Ftd>\u003Ctd align=\"right\">76.36\u003C\u002Ftd>\u003Ctd align=\"right\">0.1957\u003C\u002Ftd>\u003Ctd align=\"right\">79.79\u003C\u002Ftd>\u003Ctd align=\"right\">0.1994\u003C\u002Ftd>\u003Ctd align=\"right\">76.38\u003C\u002Ftd>\u003Ctd align=\"right\">0.2378\u003C\u002Ftd>\u003Ctd align=\"right\">66.06\u003C\u002Ftd>\u003Ctd align=\"right\">0.3426\u003C\u002Ftd>\u003C\u002Ftr>\n    \u003Ctr>\u003Ctd>\u003Cb>4RC\u003C\u002Fb>(2026)\u003C\u002Ftd>\u003Ctd align=\"right\">69.04\u003C\u002Ftd>\u003Ctd align=\"right\">0.2603\u003C\u002Ftd>\u003Ctd align=\"right\">80.27\u003C\u002Ftd>\u003Ctd align=\"right\">0.2681\u003C\u002Ftd>\u003Ctd align=\"right\">82.91\u003C\u002Ftd>\u003Ctd align=\"right\">0.1889\u003C\u002Ftd>\u003Ctd align=\"right\">84.28\u003C\u002Ftd>\u003Ctd align=\"right\">0.1766\u003C\u002Ftd>\u003C\u002Ftr>\n    \u003Ctr>\n      \u003Ctd>\u003Cb>OpenD4RT 32CLIP (Ours)\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>78.63\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>0.1811\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>66.03\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"right\">\u003Cb>0.3397\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>72.66\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"right\">\u003Cb>0.2944\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>69.93\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"right\">\u003Cb>0.2964\u003C\u002Fb>\u003C\u002Ftd>\n    \u003C\u002Ftr>\n    \u003Ctr>\n      \u003Ctd>\u003Cb>OpenD4RT 48CLIP (Ours)\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>79.60\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>0.1753\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>67.99\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"right\">\u003Cb>0.3178\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>72.48\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"right\">\u003Cb>0.2959\u003C\u002Fb>\u003C\u002Ftd>\n      \u003Ctd align=\"right\">\u003Cb>72.20\u003C\u002Fb>\u003C\u002Ftd>\u003Ctd align=\"right\">\u003Cb>0.2758\u003C\u002Fb>\u003C\u002Ftd>\n    \u003C\u002Ftr>\n  \u003C\u002Ftbody>\n\u003C\u002Ftable>\n\nTip: OpenD4RT has the strongest PStudio result in this comparison.\n\n## 🎬 Result Gallery\n\n| Case \u002F Motion | RGB + 2D Tracking | GT vs Pred 3D Tracks |\n| --- | --- | --- |\n| `softball_25`\u003Cbr>Softball swing and fast ball motion | \u003Cimg src=\"demo\u002Fsoftball_25_rgb_gt_pred_2d.gif\" width=\"360\" alt=\"Softball RGB video with GT and OpenD4RT 2D tracking overlay\"> | \u003Cimg src=\"demo\u002Fsoftball_25_gt_pred_3d.gif\" width=\"360\" alt=\"Softball GT and OpenD4RT 3D track comparison\"> |\n| `football_16`\u003Cbr>Football play with player and ball motion | \u003Cimg src=\"demo\u002Ffootball_16_rgb_gt_pred_2d.gif\" width=\"360\" alt=\"Football RGB video with GT and OpenD4RT 2D tracking overlay\"> | \u003Cimg src=\"demo\u002Ffootball_16_gt_pred_3d.gif\" width=\"360\" alt=\"Football GT and OpenD4RT 3D track comparison\"> |\n\n## 👁️ Viser Demo Visualization\n\nBuild two example Viser demo packages. Each package uses the first 64 frames:\n\n```bash\nDEMO_CASE=pstudio_mini\u002Fjuggle_5.npz OUTPUT_DIR=tmp\u002Fworldtrack_demo_juggle bash run_build_worldtrack_demo.sh\nDEMO_CASE=pstudio_mini\u002Fsoftball_25.npz OUTPUT_DIR=tmp\u002Fworldtrack_demo_softball bash run_build_worldtrack_demo.sh\n```\n\nOpen a demo package with Viser:\n\n```bash\npython vis\u002Fserve_demo_viser.py --root tmp\u002Fworldtrack_demo_juggle --port 8081\n```\n\nFor a lighter\u002Ffaster package:\n\n```bash\nDEMO_CASE=pstudio_mini\u002Fjuggle_5.npz \\\nOUTPUT_DIR=tmp\u002Fworldtrack_demo_small \\\nPOINT_GRID_COLS=32 POINT_GRID_ROWS=32 POINT_MAX_POINTS=1024 TRACK_MAX_POINTS=96 \\\nbash run_build_worldtrack_demo.sh\n```\n\nThe generated demo package contains `assets\u002Fdemo_data.json`,\n`assets\u002Finput_video.mp4`, rendered diagnostic videos, and `manifest.json`.\n\n## ✅ ToDo\n\n- [x] Release the OpenD4RT model runtime for the 32-frame 9-dataset checkpoint.\n- [x] Release WorldTrack evaluation scripts and archived metrics.\n- [x] Release Viser-based qualitative visualization tools.\n- [ ] Release complete training code.\n- [ ] Release additional checkpoints listed in the Checkpoint Zoo.\n- [ ] Release SynthVerse evaluation results.\n- [ ] Release full evaluation code for the benchmarks reported in the D4RT\n  paper and appendix.\n\n## 📄 License\n\nOpenD4RT is an unofficial implementation and is not affiliated with or endorsed\nby the original D4RT authors. The code in this repository is released under the\nApache 2.0 license; see [LICENSE](LICENSE). The D4RT paper, project page,\ndatasets, third-party assets, and upstream dependencies remain under their\nrespective licenses and terms.\n\n## 🙏 Acknowledgements\n\nThis project is built upon the D4RT paper and official project materials. We\nthank the original D4RT authors for introducing the D4RT formulation, releasing\nthe project page, and documenting the paper and appendix details that this\nimplementation follows. We also acknowledge the contributors and resources\ncredited on the official D4RT website, including colleagues who supported\nproject advice, manuscript feedback, early development, code review,\nvisualization, baseline comparisons, and data generation. We also thank the\nsplat viewer authors for the WebGL renderer used by the official D4RT\nvisualization pipeline. Please refer to the official D4RT project page for the\nfull original acknowledgements.\n","OpenD4RT 是一个非官方的 PyTorch\u002FGPU 实现，用于4D重建和跟踪。该项目基于 D4RT 模型架构，实现了从单个视频中推断深度、时空对应关系以及相机参数的功能，并提供了一个统一的查询接口来追踪3D位置。技术上，它使用了PyTorch框架，并且支持GPU加速以提高计算效率。此外，项目还提供了WorldTrack评估工具和Viser可视化工具，便于用户进行模型效果展示与分析。适合需要处理动态场景重建及物体追踪的应用场景，如自动驾驶、虚拟现实等领域的研究与开发工作。","2026-06-11 04:01:23","CREATED_QUERY"]