ACROSSVAM3.0: Resource-Bounded 4D Scene Construction and Adaptive Surface Association
Organizations: Across Physcis (Beijing) Technology Co., Ltd
Abstract
Combining video generation, dynamic reconstruction and monocular SLAM requires consistent coordinates, explicit handling of missing outputs and predictable memory use. We present AcrossVAM3.0, an offline system that converts text, images or monocular RGB video into a versioned point representation. Frozen components run sequentially on one GPU; map corrections update cameras, points and trajectories together. The coordination layer adds no neural parameters. Three fresh input-mode runs completed with a maximum sampled task-process-tree GPU footprint of 17.14 GB. In a frozen 16-scene video suite, every scene yielded a 4D reconstruction and 15 yielded a non-empty SLAM map. Strict cross-window matching reduced local surface error but lost long-range track coverage on TAP-Vid-DAVIS. We therefore developed multi-hypothesis adaptive surface chaining (MHAC), which retains several candidate matches and an explicit missing state. After selection on disjoint development data, MHAC improved Average Jaccard over a position-only baseline by 0.1394 on a frozen ten-video holdout (95% bootstrap interval [0.1074,0.1689]). Across 265 RoboTAP videos, the mean paired gain was 0.0610 ([0.0510,0.0715]), with 215 wins, one tie and 49 losses. These results support resource-bounded scene construction and an external tracking gain under the tested protocol. They do not establish metric scale, semantic identity or high-fidelity novel-view rendering.
Figures & tables
| Evidence block | Independent units | Status | Resource peak | Claim boundary |
| Three-input acceptance | 9 runs + 3 fresh final runs | Complete | 17.14 GB | Executability of fixed short-scene profile; generation quality separate |
| Frozen real-video suite | 16 scenes | Complete | 16.27 GB | Point 4D reconstruction, RGB-only SLAM runtime, scene-level diagnostics |
| P4 TAP-Vid-DAVIS | 30 videos | Complete; primary hypothesis failed | 26.54 GB | External falsification of strict three-frame chain |
| P5 RGB-Stacking calibration | 30 videos | Complete; development | 16.27 GB | Candidate design only |
| P5 selection validation | 10 videos | Complete; development | 16.27 GB | Deterministic candidate selection only |
| P5 development holdout | 10 videos | Complete; gate passed | 16.27 GB | One-shot authorization of final external test |
| Input | Frames | Submaps | Time (s) | GPU (GB) | Execution |
| Text | 17 | 2 | 262.32 | 17.12 | Pass |
| Image | 17 | 2 | 259.59 | 17.14 | Pass |
| Observed RGB | 60 | 5 | 155.03 | 16.27 | Pass |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Claim | Artifact | Evidence class |
| Three-input final acceptance | reports/phase3/d30_final_acceptance.json | Final engineering acceptance |
| Frozen 16-scene metrics | reports/phase3/d25_statistics.json and case CSV | Frozen test; descriptive inference |
| P4 external negative result | reports/phase4/p4_tapvid_aggregate.json | Preregistered external test |
| P5 candidate selection | reports/phase5/p5_mhac_selection_validation.json | Development selection |
| P5 frozen holdout | reports/phase5/p5_mhac_development_holdout.json | One-shot development holdout |
| P5 RoboTAP external result | reports/phase5/p5_robotap_external_aggregate.json | Frozen external test |
Dynamic Scene Reconstruction
Dynamic scene reconstruction builds time-varying 3D models of moving objects and environments from video or multi-view images.
See all 311 papers