simple_visual_slam

Reference Keyframe Policy Summary

This is the GitHub-friendly landing page for the reference-keyframe experiment track. It summarizes the current conclusion and links to the detailed experiment artifacts.

Current Status

• Curated corpus top accuracy: score (0.929), pipeline (0.929).
• Bounded real-trace replay winner: score with mean APE 0.000.
• Full repeat-2 replay winner: score with mean/std 0.093±0.080.
• Runtime default is still heuristic because no single policy dominates every mode under repeat replay.

Mode Snapshot

• Single-run bounded replay mode winners: ``.
• Repeat-2 bounded replay mode winners: depth=heuristic, depth_accel=score, mono=score.
• Room-only hotspot winners: depth_accel=score, mono=pipeline.

Read Next

• Detailed experiment tables
• Current decisions and adoption criteria
• Minimal interface that survived the experiments

Reproduce

bash scripts/eval_reference_policies.sh --repeat 1
bash scripts/eval_reference_policies.sh --repeat 2 --output eval_results/reference_keyframe_policy/real_trace_metrics_repeat2.csv
bash scripts/eval_reference_policies.sh --repeat 2 --corpus experiments/reference_keyframe/room_focus_corpus.tsv --output eval_results/reference_keyframe_policy/room_focus_repeat2.csv
./scripts/update_reference_policy_docs.py

This site is open source. Improve this page.