Learning Proposes, Geometry Disposes: A Modular Framework for Efficient Spatial Reasoning

1. Problem Statement

The paper addresses a fundamental ambiguity in modern spatial perception systems: how to integrate learning-based methods with classical geometric pipelines.

• The Conflict: Traditional visual odometry (VO) and SLAM rely on explicit geometric optimization (e.g., Bundle Adjustment, ICP) to enforce physical consistency. However, these methods are sensitive to initialization and data association errors. Conversely, recent learning-based approaches (e.g., VGGT, DROID-SLAM) demonstrate strong representational capacity for pose and depth estimation but often lack explicit mechanisms to enforce geometric consistency, leading to instability under large baselines, distribution shifts, or dynamic scenes.
• The Core Question: Should learning components directly replace geometric estimation, or should they serve as intermediate modules? Current hybrid approaches often treat learning outputs as soft priors or final estimates without a clear separation between hypothesis generation and validation, leading to error propagation.

2. Methodology: "Learning Proposes, Geometry Disposes"

The authors propose a modular, end-to-end framework that explicitly separates the roles of learning and geometry:

• Learning Proposes: Data-driven models generate geometric hypotheses (pose and depth) but are treated as suggestions, not commitments. These proposals encode statistical regularities but are not required to strictly satisfy rigid-body constraints.
• Geometry Disposes: A classical geometric backend acts as a decision boundary. It validates, refines, or rejects the learning proposals based on physical and geometric consistency (e.g., multi-view consistency, reprojection error).

Concrete Implementation

The authors instantiate this framework using the following components:

1. Learning Module (Proposal Generator): They utilize VGGT (Wang et al., 2025), a state-of-the-art model that predicts dense depth and relative camera pose from monocular RGB inputs. Crucially, VGGT is used without fine-tuning, and its outputs are treated purely as initial hypotheses.
2. Geometric Disposer (Decision Maker): A classical Point-to-Plane RGB-D Iterative Closest Point (ICP) algorithm serves as the backend.
   • Process: The ICP is initialized with the pose proposal from VGGT.
   • Verification: The ICP optimizes the transformation to minimize alignment error. If the optimization fails to converge or the final residual exceeds a threshold, the proposal is rejected.
   • Refinement: If the proposal passes, the ICP refines the pose, ensuring the final output adheres to rigid-body geometry.
3. Execution: The system operates in an online, incremental manner, processing frames sequentially to expose failure modes like drift accumulation.

3. Key Contributions

1. Conceptual Framework: The paper introduces a clear paradigm shift: "Learning proposes, Geometry disposes." It argues that learning should explore the solution space (hypothesis generation), while geometry should act as the arbiter of physical feasibility (decision making).
2. System Design: A modular architecture where learning outputs are strictly decoupled from the final decision authority. This prevents "hallucinations" from learning models from dominating the optimization process.
3. Empirical Validation: A systematic evaluation on the TUM RGB-D benchmark comparing three configurations:
   • Learning-only (direct pose prediction).
   • Geometry-only (ICP with identity initialization).
   • Hybrid (Learning proposes + Geometry disposes).

4. Experimental Results

Experiments were conducted on the TUM RGB-D dataset across varying motion magnitudes (frame strides) and scene dynamics (static vs. dynamic).

• Reliability of Learning Proposals: Learning-based pose proposals alone were found to be unreliable. In many cases, they performed worse than a simple identity initialization, particularly at larger frame strides (high motion).
• The Power of Geometric Disposal:
  • Initialization Independence: After geometric disposal, the final performance became largely insensitive to the quality of the initial pose proposal. Whether initialized with a VGGT prediction or an identity matrix, the final Relative Pose Error (RPE) was nearly identical.
  • Error Correction: The geometric backend effectively corrected erroneous learning hypotheses or rejected them entirely, preventing error accumulation.
  • Robustness: The hybrid system maintained consistent performance across static and dynamic scenes, whereas learning-only methods degraded significantly in dynamic environments.
• Depth Alignment: The study highlighted that learning-proposed geometry must be geometrically aligned with camera intrinsics; misalignment can degrade performance, further emphasizing the need for a geometric verification stage.

5. Significance and Implications

• Geometry as a Safeguard: The paper demonstrates that geometry is not merely a refinement step but an essential arbiter. It validates learning-based observations against physical laws (rigid-body constraints).
• Redefining the Role of Learning: The authors suggest a shift in how learning modules should be designed for spatial reasoning. Instead of optimizing for standalone accuracy, learning models should be designed to generate diverse, informative hypotheses that facilitate downstream geometric verification.
• System Stability: By treating learning outputs as proposals rather than commitments, the framework achieves robustness that is difficult to attain with end-to-end learning or pure geometric methods alone.
• Future Directions: The work opens avenues for integrating uncertainty-aware learning models and extending the "disposal" mechanism to include long-term loop closures and global optimization, while maintaining the core principle of geometric validation.

In conclusion, the paper argues that for robust spatial perception, learning should explore, and geometry should decide. This modular approach ensures that systems remain data-adaptive while strictly adhering to physical consistency.