Research map / 3D vision / Gaussian splatting
MAC-Ego3D: Multi-Agent Gaussian Consensus for Real-Time Collaborative Ego-Motion and Photorealistic 3D Reconstruction
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 MAC-Ego3D, a real-time multi-agent RGB-D SLAM system that uses 3D Gaussian splats with intra- and inter-agent Gaussian consensus for ego-motion tracking and photorealistic dense mapping.
Problem
Real-time collaborative ego-motion estimation and high-fidelity dense 3D reconstruction across multiple agents, where sparse SLAM maps lack detail and dense neural multi-agent methods have high latency.
Approach
Represents each agent's local map with 3D Gaussian splats and aligns agents by selecting, associating, aligning and jointly optimizing co-visible Gaussians (Intra- and Inter-Agent Gaussian Consensus), with asynchronous parallel execution per agent.
Key results
- Average ATE RMSE of 0.14 cm on Multi-agent Replica versus 1.23 cm for CP-SLAM.
- Average RGB PSNR of 40.04 dB on Multi-agent Replica versus 29.33 dB for CP-SLAM; on 7-Scenes 21.02 dB versus 17.20 dB.
- Tracking 0.02 s and mapping 0.05 s per frame on loop sequences, versus CP-SLAM's 0.30 s tracking and 10.10 s mapping.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent SLAM · Multi-agent Replica, Average | ATE RMSE [cm]: 1.23 -> 0.14 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent dense mapping · Multi-agent Replica, Average | RGB PSNR [dB]: 29.33 -> 40.04 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent dense mapping · Multi-agent Replica, Average | SSIM: 0.843 -> 0.977 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent dense mapping · Multi-agent Replica, Average | LPIPS: 0.314 -> 0.050 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent dense mapping · Multi-agent Replica, Average | Depth L1 Loss [mm]: 1.40 -> 0.74 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent SLAM · 7-Scenes, Average | RGB PSNR [dB]: 17.20 -> 21.02 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent SLAM · 7-Scenes, Average | SSIM: 0.592 -> 0.774 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent SLAM · 7-Scenes, Average | LPIPS: 0.546 -> 0.283 | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent SLAM · Multi-agent Replica, loop sequences | Tracking time per frame: 0.30s -> 0.02s | Helped |
| Use MAC-Ego3D instead of CP-SLAM | Multi-agent SLAM · Multi-agent Replica, loop sequences | Mapping time per frame: 10.10s -> 0.05s | Helped |
| Add depth-noise robustness through MAC-Ego3D | Multi-agent SLAM · Perturbed Multi-agent Replica, depth noise | ATE [cm]: 1.78 cm -> 0.27 cm | Helped |
| Add RGB-noise robustness through MAC-Ego3D | Multi-agent SLAM · Perturbed Multi-agent Replica, RGB corruption | ATE [cm]: 1.08 cm -> 0.14 cm | Helped |
Limitations
- Does not address associating patterns across multiple rooms at scale.
- No communication bandwidth constraints imposed; communication-constrained settings left for future work.
- Only sequential multi-agent collaboration; parallel collaboration deferred.
Experimental setup
Datasets: Multi-agent Replica 7-Scenes
Benchmarks: Multi-agent Replica 7-Scenes
Metrics: ATE RMSE PSNR SSIM LPIPS depth L1 loss tracking and mapping time per frame
Baselines: CCM-SLAM ORB-SLAM3 Swarm-SLAM CP-SLAM NICE-SLAM Vox-Fusion
Builds on: 3D Gaussian Splatting G-ICP CP-SLAM
Closest papers by meaning
- Scalable Benchmarking and Robust Learning for Noise-Free Ego-Motion and 3D Reconstruction from Noisy Video (ICLR 2025 · similarity 0.88)
- Reinventing Multi-Agent Collaboration through Gaussian-Image Synergy in Diffusion Policies (NeurIPS 2025 · similarity 0.88)
- FRUC: Feedforward Dynamic Scene Reconstruction from Uncalibrated Collaborative Driving Views (NeurIPS 2026 · similarity 0.87)
- MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM (CVPR 2025 · similarity 0.87)
- One World, Dual Timeline: Decoupled Spatio-Temporal Gaussian Scene Graph for 4D Cooperative Driving Reconstruction (NeurIPS 2026 · similarity 0.86)
- PF3plat: Pose-Free Feed-Forward 3D Gaussian Splatting for Novel View Synthesis (ICML 2025 · similarity 0.85)
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.