simple_visual_slam

Current Minimal Interface

Reference Keyframe Policy

struct ReferenceKeyframePolicyInput {
    int tracked_features;
    int detected_keypoints;
    int candidate_landmarks;
    int frames_since_reference;
    int lost_frames;
    bool has_depth;
    bool has_accel;
};

struct ReferenceKeyframeDecision {
    ReferenceKeyframeAction action;
    double confidence;
    std::string reason;
};

Why These Fields Survived

• tracked_features: every policy needs current tracking density.
• detected_keypoints: separates sparse-image failure from healthy mono frames.
• candidate_landmarks: lets policies reason about how much map support the candidate keyframe brings.
• frames_since_reference: makes keyframe aging explicit instead of hidden inside each implementation.
• lost_frames: keeps recovery pressure observable to the policy instead of coupling it to Tracking internals.
• has_depth: the primary modality flag that every implementation already relied on.
• has_accel: survived once the room follow-up showed depth and depth_accel windows do not behave the same under replay.

Why Other Fields Were Removed

• Raw Frame, Keyframe, and Map pointers were excluded because they destroy comparability and make policy evaluation depend on global SLAM state.
• Optimizer, loop-closing, and descriptor state were excluded because none of the current variants needed them to make a comparable decision.

This site is open source. Improve this page.