iSAM2 KITTI Pipeline (Unified Methods)

This repo provides a single flag-based Stereo VO + iSAM2 pipeline for KITTI odometry.

Methods

Use run_isam2_kitti.py (single sequence) or run_all_gt_sequences.py (batch) with flags:

• base: odometry-only iSAM2 (default)
• loop-closure: add geometric loop-closure factors (--enable-loop-closure)
• confidence-v1: confidence-weighted odometry (--enable-confidence-weighting)
• confidence-v2: stronger confidence weighting + robust kernels (--enable-confidence-v2)
• info-weighted: information-weighted odometry using inlier count, match count, and reprojection error (--enable-info-weighting)
• degeneracy-aware+loop: observability-aware odometry noise scaling with loop closure (--degeneracy-aware --enable-loop-closure)

Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

KITTI Dataset Layout

Expected --dataset-root structure:

<KITTI_DATASET_ROOT>/
	poses/
		00.txt
		01.txt
		...
	sequences/
		00/
			image_0/
			image_1/
			calib.txt
			times.txt
		01/
			...

Single Sequence Examples

Use --seq 00 (or any sequence id) and choose one method configuration.

1) Base

python3 run_isam2_kitti.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--seq 00 \
	--fallback-no-motion \
	--output output/experiments/base/seq00_allframes/poses_est_00.txt \
	--metrics-out output/experiments/base/seq00_allframes/metrics_00.json

2) Loop Closure

python3 run_isam2_kitti.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--seq 00 \
	--fallback-no-motion \
	--enable-loop-closure \
	--output output/experiments/loop_closure/seq00_allframes/poses_est_00.txt \
	--metrics-out output/experiments/loop_closure/seq00_allframes/metrics_00.json

3) Confidence v1

python3 run_isam2_kitti.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--seq 00 \
	--fallback-no-motion \
	--enable-confidence-weighting \
	--confidence-floor 0.08 \
	--confidence-min-scale 0.85 \
	--confidence-max-scale 4.0 \
	--output output/experiments/confidence_v1/seq00_allframes/poses_est_00.txt \
	--metrics-out output/experiments/confidence_v1/seq00_allframes/metrics_00.json

4) Confidence v2

python3 run_isam2_kitti.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--seq 00 \
	--fallback-no-motion \
	--enable-confidence-v2 \
	--enable-loop-closure \
	--confidence-v2-robust-kernel huber \
	--confidence-v2-robust-k 1.5 \
	--output output/experiments/loop_closure_confidence_v2/seq00_allframes/poses_est_00.txt \
	--metrics-out output/experiments/loop_closure_confidence_v2/seq00_allframes/metrics_00.json

5) Info-Weighted

python3 run_isam2_kitti.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--seq 00 \
	--fallback-no-motion \
	--enable-loop-closure \
	--enable-info-weighting \
	--info-min-confidence 0.1 \
	--info-max-confidence 1.0 \
	--output output/experiments/info_weighted_loop/seq00_allframes/poses_est_00.txt \
	--metrics-out output/experiments/info_weighted_loop/seq00_allframes/metrics_00.json

6) Degeneracy-Aware + Loop

python3 run_isam2_kitti.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--seq 00 \
	--fallback-no-motion \
	--enable-loop-closure \
	--degeneracy-aware \
	--degeneracy-cond-ref 800 \
	--degeneracy-noise-max-mult 25 \
	--output output/experiments/degeneracy_loop/seq00_allframes/poses_est_00.txt \
	--metrics-out output/experiments/degeneracy_loop/seq00_allframes/metrics_00.json

Batch Runs (All GT Sequences)

Base:

python3 run_all_gt_sequences.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--fallback-no-motion \
	--output-dir output/experiments/base/batch_gt

Loop closure:

python3 run_all_gt_sequences.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--fallback-no-motion \
	--enable-loop-closure \
	--output-dir output/experiments/loop_closure/batch_gt

Confidence v2 + loop:

python3 run_all_gt_sequences.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--fallback-no-motion \
	--enable-loop-closure \
	--enable-confidence-v2 \
	--output-dir output/experiments/loop_closure_confidence_v2/batch_gt

Info-weighted + loop:

python3 run_all_gt_sequences.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--fallback-no-motion \
	--enable-loop-closure \
	--enable-info-weighting \
	--output-dir output/experiments/info_weighted_loop/batch_gt

Degeneracy-aware + loop:

python3 run_all_gt_sequences.py \
	--dataset-root <KITTI_DATASET_ROOT> \
	--fallback-no-motion \
	--enable-loop-closure \
	--degeneracy-aware \
	--output-dir output/experiments/degeneracy_loop/batch_gt

Flag Compatibility Notes

• --enable-confidence-v2 overrides --enable-confidence-weighting behavior.
• --enable-info-weighting cannot be combined with confidence-v1 or confidence-v2.
• --degeneracy-aware currently:
  • can be used with or without loop closure
  • cannot be combined with confidence-v1, confidence-v2, or info-weighting

Output Layout

This repo currently has two output organizations:

• Legacy per-method folders (existing historical outputs under output/)
• Consolidated folders:
  • output/experiments/
  • output/comparisons/
  • output/manifests/

The consolidation copy manifest is saved at:

• output/manifests/output_migration_manifest.json

Plotting and Evaluation

• Batch metric plots: plot_batch_metrics.py
• Estimate-vs-GT comparisons: plot_estimate_vs_gt.py
• Base-vs-loop comparison: plot_base_vs_loop.py
• Metric utilities: eval/compute_kitti_metrics.py, eval/loop_closure_gain.py

Use --help on each script for full options.

Fallback Branch Runs (If Main Outputs Do Not Reproduce)

If the unified main branch does not reproduce expected outputs, run the original method groups on their source branches separately:

• loop_vs_loopconf: methods M1-M6
• degeneracy2.0: methods M7-M8
• mohammad/information-based-weighting: method M9

Suggested workflow:

git fetch origin

# M1-M6
git checkout loop_vs_loopconf

# M7-M8
git checkout degeneracy2.0

# M9
git checkout mohammad/information-based-weighting

# return to unified implementation
git checkout main

Run each branch's method scripts/configs independently so the nine implementations are executed in their original branch context.