The pose of the reference camera of a rig, expressed in the frame of the anchor’s reference camera, is estimated in one feed-forward pass. The anchor advances with the stream.
Related Work
Multi-Camera Odometry
Multiple cameras widen the field of view and add redundancy when one view degrades. The generalized camera model treats a calibrated rig as one imaging device, supporting odometry with little or no view overlap and keyframe SLAM with arbitrary rig layouts. These systems recover motion by iterative optimization over explicit geometric residuals.
Streaming models retain history for single-camera inputs, while rig-aware models jointly perceive calibrated views without temporal state. To our knowledge, prior work has not adapted a frozen multi-view 3D foundation model to causal odometry with calibrated multi-camera rigs. StreamRig fills this gap.
Method
StreamRig is a causal streaming odometer for calibrated multi-camera rigs. It keeps a multi-view 3D foundation model frozen as the perception front-end and adds three lightweight modules that turn per-rig perception into a causal state and read odometry from it. Only these modules are trained, with relative poses as their sole supervision.
Streaming Inference and Re-Anchoring
At each step, the backbone perceives the new rig once, the Rig-Resampler compresses it, and its tagged tokens join the active window. The keys and values of every earlier rig are cached in all eight layers of the bridge. One pass of the bridge over the new rig and its anchor snapshot, followed by the pose head, returns the relative pose estimate for the current rig with respect to the anchor.
Experimental Setup
Protocol
Every method runs causally over each complete sequence, anchored to its first pose, and composes its trajectory from its own predictions. Except for the MA-AB oracle, ground-truth poses are used only for evaluation.
Each method re-anchors in blocks of a specified number of sampled rigs or frames. Thus, the pose estimate is formed as a pairwise chain, and a larger block size retains a longer state.
Conclusion
The results of StreamRig support joint rig perception and temporal history as complementary sources of odometry accuracy. The Rig-Resampler and CausalBridge make this combination affordable, while group relocalization provides the alignment initialization that limited rig data alone does not supply. The re-anchoring study shows that longer training windows extend the usable anchor range. Effective streaming combines compact geometric observations with an appropriate re-anchoring interval.
Future Work
Exploratory trials on NCLT suggest that compressed rig tokens may be useful for same-session revisit retrieval, but this capability is not quantitatively evaluated here, and acceptance thresholds still require per-session tuning. Future work will develop transferable loop triggers and integrate revisits into the streaming state. It will also jointly refine poses and dense geometry so that the accumulated map constrains subsequent estimates.
Frequently Asked Questions
What problem does StreamRig address? It addresses causal odometry for calibrated multi-camera rigs by combining rig perception with temporal history.
Does StreamRig train its multi-view 3D foundation model? No. The foundation model remains frozen, and only the three added lightweight modules are trained.
How does StreamRig process each new rig? The backbone perceives it, the Rig-Resampler compresses it, and its tagged tokens join the active window. Cached keys and values provide the earlier rigs’ state to the bridge.
What does the paper identify as future work? It proposes developing transferable loop triggers, integrating revisits into the streaming state, and jointly refining poses and dense geometry.