Research map / 3D vision / 3D generation
KV-Tracker: Real-Time Pose Tracking with Transformers
Marwan Taher, Ignacio Alzugaray, Kirill Mazur, Xin Kong, Andrew Davison
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| cache keys/values from keyframes and track single query frames instead of recomputing all-to-all attention (ablation) | pi3 multi-view model · Runtime analysis (308x308) | FPS: 30 FPS up to 50 frames vs quadratic slowdown | Improved |
| use KV-cache tracking instead of streaming memory (TTT3R) | KV-Tracker · TUM RGB-D | ATE RMSE (m): 0.132 -> 0.098 | Improved |
| use KV-cache tracking instead of TTT3R | KV-Tracker · 7-Scenes | ATE RMSE (m): 0.143 -> 0.059 | Improved |
| KV-Tracker vs sparse DPVO | KV-Tracker · TUM RGB-D | ATE RMSE (m): 0.095 -> 0.098 | No clear effect |
| use KV-Tracker instead of TTT3R at lower resolution | KV-Tracker · ARCTIC | ATE RMSE (m): 0.303 -> 0.228 | Improved |
| raise input resolution from 308 to 518 (2D bbox) (ablation) | KV-Tracker · OnePose | recall 3cm-3deg (%): 52.8 -> 69.3 | Improved |
| raise input resolution from 308 to 518 (2D bbox) (ablation) | KV-Tracker · OnePose LowTexture | recall 5cm-5deg (%): 83.3 -> 94.4 | Improved |
| use dilated 2D bboxes instead of segmentation masks (ablation) | KV-Tracker @518 · OnePose and OnePose LowTexture | recall (%): 92.1 -> 92.9 (OnePose 5cm); 10.7 -> 5.3 (OnePose 1cm) | Mixed |
| online map vs baselines' offline 3D reconstruction (strict threshold) | KV-Tracker · OnePose | recall 1cm-1deg (%): 51.1 -> 10.7 | Hurt |
| online KV-Tracker vs OnePose++ with offline maps (coarse threshold) | KV-Tracker @518 2D bbox · OnePose LowTexture | recall 5cm-5deg (%): 72.1 -> 94.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- TTT3R — TTT3R: 3D Reconstruction as Test-Time Training (ICLR 2026) · better on 5 of 5 shared results
- OnePose — OnePose: One-Shot Object Pose Estimation Without CAD Models (CVPR 2022) · better on 4 of 5 shared results
- Point3R — Point3R: Streaming 3D Reconstruction with Explicit Spatial Pointer Memory (NeurIPS 2025) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Seeing Behind Objects for 3D Multi-Object Tracking in RGB-D Sequences (CVPR 2021 · similarity 0.89)
- KPA-Tracker: Towards Robust and Real-Time Category-Level Articulated Object 6D Pose Tracking (AAAI 2024 · similarity 0.88)
- Free-Moving Object Reconstruction and Pose Estimation with Virtual Camera (AAAI 2025 · similarity 0.87)
- TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking (NeurIPS 2026 · similarity 0.87)
- ProxyPose: 6-DoF Pose Tracking via Video-to-Video Translation (NeurIPS 2026 · similarity 0.86)
- TAPIP3D: Tracking Any Point in Persistent 3D Geometry (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
