DAC-MACE: Context-Conditioned Expert Routing for Compact Scene Coordinate Regression in Camera Relocalization

Dingying Fan, Mingkai Liu, Haohua Que, Haojia Gao, Xianliang Huang, Qian Zhang, Xueqian Wang, Fei QiaoarXiv 2026 · arXiv

TL;DR

DAC-MACE is a new approach to camera relocalization that solves the "repeated facade" problem in large-scale scenes. By adding global context and self-attention to a Mixture-of-Experts (MoE) architecture, it achieves state-of-the-art accuracy (13.0 cm mean error on Cambridge Landmarks) while maintaining a tiny 10 MB memory footprint.

What Problem Does This Paper Address?

Imagine a mobile robot trying to figure out exactly where it is in a large city square. It looks at a window, but that window looks identical to a hundred others on the same building. This is perceptual ambiguity.

Current "Scene Coordinate Regression" (SCR) methods try to map every pixel in an image to a 3D coordinate. To handle large areas, they often use a Mixture-of-Experts (MoE), where different "expert" networks handle different parts of the scene. However, the "router"—the part that decides which expert to use—usually only looks at a small, local patch of the image. If two patches look the same, the router gets confused, sends the data to the wrong expert, and the resulting 3D coordinate is wrong.

Methods

The authors introduce DAC-MACE (Dynamic Attention-driven Context MACE), which gives the router a "big picture" view of the scene.

The DAC-MACE pipeline: from a frozen encoder to context-conditioned routing and a dual-threshold pose estimator.

The system builds on the ACE (Accelerated Coordinate Encoding) framework but adds three key "contextual" modules:

  1. Residual Context Adapter (RCA): A per-scene adapter that reweights feature channels and locations to emphasize what's important for a specific environment.
  2. Global Context Encoder (GCE): This is the heavy lifter. It uses multi-scale self-attention to allow every feature in the image to "talk" to every other feature. This helps the network distinguish between two identical windows by looking at their surroundings.
  3. Image-level Pooled Descriptor: A final summary of the entire image is fed into the router to provide a global "anchor."

Finally, they use a Dual-Threshold RANSAC for pose estimation. Instead of picking one fixed error tolerance, they run two searches (one strict, one loose) and use a clever selection rule to pick the most reliable camera pose.

Datasets

The authors evaluated their method on two standard benchmarks:

  • Cambridge Landmarks: Five large-scale outdoor scenes (up to 8,000 m²). This is where the "repeated facade" problem is most prominent.
  • 7-Scenes: Seven small-scale indoor scenes used to test if the context modules help even when a single expert (no MoE) is used.

Key Insights and Ideas

  • Context is the Cure for Ambiguity: The "Aha!" moment is that local descriptors are too myopic. By using self-attention, the network can see that "this window is near the roof" while "that identical window is near the ground," allowing the router to pick the correct expert.
  • Efficiency via Frozen Encoders: Like its predecessor ACE, DAC-MACE keeps the main feature encoder frozen. This allows for incredibly fast "mapping" (training on a new scene) in just 35 minutes, compared to hours or days for older methods.
  • Soft Routing: Instead of picking just one expert (hard routing), DAC-MACE uses "soft" weights to blend the outputs of multiple experts, leading to smoother and more accurate coordinate predictions.
Qualitative diagnostic showing how GCE concentrates the feature response, reducing ambiguity on a repetitive facade.

Key Results

  • Top Accuracy: On Cambridge Landmarks, DAC-MACE achieved a mean translation error of 13.0 cm, outperforming other RGB-only SCR methods like GLACE (13.6 cm) and MACE (14.4 cm).
  • Big Wins on Big Scenes: On the "Great Court" (the largest and most difficult scene), it reduced the error from 24 cm to 19 cm.
  • Compactness: Despite the high accuracy, the map size is only 10 MB. As shown in the figure below, it offers a much better accuracy-to-memory trade-off than traditional feature-matching methods that require hundreds of megabytes.
Fig. 2. Memory–accuracy trade-off on Cambridge Landmarks: map size versus the mean of per-scene median translation errors (values of Table II). DAC- MACE has the lowest error among the RGB-only SCR methods. Plotted method/value pairs: AS (200 MB, 13.8 cm), hLoc (800, 10.8), PixLoc (600, 15.0), DSAC* with depth (28, 20.6), SRC with depth (40, 41.6), SANet with depth (260, 83.6), RGB-only DSAC* (28, 18.6), DSAC* Tiny (4, 45.0), ACE (4, 25.0), Poker (16, 17.0), GLACE (13, 13.6), MACE (5, 14.4), and DAC-MACE (10, 13.0).

Limitations and Future Work

While highly effective, DAC-MACE has a few hurdles:

  • Computational Cost: The self-attention mechanism in the GCE is computationally expensive (adding about 98 GFLOPs). This might be a challenge for low-power embedded devices.
  • Repetitive Geometry Failure: On the "Stairs" scene in the indoor dataset, the error actually increased. The authors note that when geometry is truly identical (like a flight of stairs), content-based attention without positional info can actually make things more confusing.
  • Future Directions: The authors suggest exploring "lighter" attention mechanisms (like windowed or linear attention) to make the system more suitable for real-time use on mobile hardware.

Summary updated

Abstract

A camera becomes an absolute pose sensor when its images can be relocalized against a prebuilt map of the scene. On mobile robots and augmented-reality devices, that map must be small and quick to build. Scene coordinate regression (SCR) stores the scene in a compact network that predicts a 3D coordinate for every pixel, and mixture-of-experts (MoE) variants extend its capacity to larger outdoor scenes. In existing MoE-based SCR, however, the router assigns each pixel to an expert from the same local descriptor that the experts regress from, so repeated facades and windows at different locations receive similar routing inputs. We present DAC-MACE, which conditions soft expert routing on recalibrated and contextualized features. A residual context adapter (RCA) reweights the channels and locations of frozen encoder features, a global context encoder (GCE) mixes them with multi-scale self-attention, and an image-level pooled descriptor is added to the router input. A dual-threshold RANSAC--Levenberg--Marquardt estimator recovers the pose. On Cambridge Landmarks, DAC-MACE attains a mean per-scene median translation error of 13.0 cm over three seeds, the lowest among the RGB-only SCR methods compared (GLACE 13.6 cm, MACE 14.4 cm), and reduces the error on the largest scene from 24 to 19 cm relative to MACE. In a cumulative study, the MoE head, GCE, RCA, and the dual-threshold estimator lower the mean error from 17.0 cm to 14.4, 14.0, 13.6, and 13.0 cm. On 7-Scenes, adding GCE to a single ACE head raises Recall@(5 cm, 5$^\circ$) from 86.56% to 87.60%.
BibTeX
@misc{fan2026dac,
  author = {Dingying Fan and Mingkai Liu and Haohua Que and Haojia Gao and Xianliang Huang and Qian Zhang and Xueqian Wang and Fei Qiao},
  title = {{DAC-MACE: Context-Conditioned Expert Routing for Compact Scene Coordinate Regression in Camera Relocalization}},
  year = {2026},
  eprint = {2611.10263},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10263},
}
Figure