Research map / Detection and segmentation / Tracking
CTO-SLAM: Contour Tracking for Object-Level Robust 4D SLAM
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CTO-SLAM, an object-level 4D SLAM system that tracks contour keypoints of dynamic objects with pixel tracking, initializes object poses by sampling yaw, and refines object and camera poses in a contour-tracking bundle adjustment.
Problem
Robust long-term spatiotemporal data association for tracking dynamic objects and estimating their motion jointly with camera ego-motion in monocular and RGB-D SLAM for dynamic outdoor scenes.
Approach
Track contour keypoints on dynamic objects using pixel tracking (TAP-style) instead of appearance features, initialize object poses by ground fitting and sampled yaw cuboid fitting, and jointly optimize camera, object, static and contour points in a factor graph with the sampling-based poses weighted higher.
Key results
- On KITTI Tracking ego-motion, CTO-SLAM reaches average ATE 1.05 m versus 2.33 m for ORB-SLAM2, 1.45 m for DynaSLAM and 1.54 m for DynaSLAM II.
- Dynamic object motion ATE averages 0.64 m for CTO-SLAM versus 0.95 m for DynaSLAM II on KITTI Tracking, a reported 33% improvement.
- Ablation on VKITTI shows camera ATE improves substantially as static structure is enriched (Full vs BK vs IN settings), supporting the claim that inaccurate object motion degrades ego-motion when jointly optimized.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ORB-SLAM2 with CTO-SLAM | camera ego-motion estimation · KITTI Tracking, average | ATE (m): 2.33 -> 1.05 | Helped |
| replace DynaSLAM with CTO-SLAM | camera ego-motion estimation · KITTI Tracking, average | ATE (m): 1.45 -> 1.05 | Helped |
| replace DynaSLAM II with CTO-SLAM | camera ego-motion estimation · KITTI Tracking, average | ATE (m): 1.54 -> 1.05 | Helped |
| replace DynaSLAM II with CTO-SLAM | dynamic object motion estimation · KITTI Tracking, average | ATE: 0.95 -> 0.64 | Helped |
| use static-background-only images instead of full images (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 01 | ATE mean (m): 0.11860 -> 0.05340 | Helped |
| inpaint dynamic foreground objects into static backgrounds (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 01 | ATE mean (m): 0.11860 -> 0.01600 | Helped |
| use static-background-only images instead of full images (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 02 | ATE mean (m): 0.00700 -> 0.00945 | Hurt |
| inpaint dynamic foreground objects into static backgrounds (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 02 | ATE mean (m): 0.00700 -> 0.00685 | Helped |
| use static-background-only images instead of full images (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 06 | ATE mean (m): 0.03660 -> 0.03231 | Helped |
| inpaint dynamic foreground objects into static backgrounds (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 06 | ATE mean (m): 0.03660 -> 0.00058 | Helped |
| use static-background-only images instead of full images (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 18 | ATE mean (m): 0.08838 -> 0.09625 | Hurt |
| inpaint dynamic foreground objects into static backgrounds (ablation) | CTO-SLAM camera ego-motion estimation · VKITTI, sequence 18 | ATE mean (m): 0.08838 -> 0.08139 | Helped |
Surprising findings
- Jointly optimizing dynamic object motion with bundle adjustment can degrade camera ego-motion accuracy when static structure is scarce, a limitation the authors identify.
Limitations
- Object motion accuracy depends on the number of visible contour keypoints; occlusion reduces points and can cause tracking failure.
- Joint camera-object bundle adjustment hurts ego-motion when static structure is limited.
- Relies on instance segmentation and monocular depth estimation (depthNet) as preprocessing.
Experimental setup
Datasets: KITTI Tracking Virtual KITTI (VKITTI)
Benchmarks: KITTI Tracking VKITTI
Metrics: RPE ATE
Baselines: ORB-SLAM2 DynaSLAM DynaSLAM II CubeSLAM SIFT ORB
Builds on: ORB-SLAM2 DROID-SLAM TAP-Vid VDO-SLAM DynaSLAM
Closest papers by meaning
- 4DTAM: Non-Rigid Tracking and Mapping via Dynamic Surface Gaussians (CVPR 2025 · similarity 0.87)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · similarity 0.85)
- Photo-SLAM: Real-time Simultaneous Localization and Photorealistic Mapping for Monocular Stereo and RGB-D Cameras (CVPR 2024 · similarity 0.84)
- VTGaussian-SLAM: RGBD SLAM for Large Scale Scenes with Splatting View-Tied 3D Gaussians (ICML 2025 · similarity 0.84)
- Loopy-SLAM: Dense Neural SLAM with Loop Closures (CVPR 2024 · similarity 0.83)
- WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments (CVPR 2025 · similarity 0.83)
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.