SLAM Learning

Map repair?

Dynamic robust mapping · Semantic mapping and localization

English | 中文

Directions · Questions · Related works · Results

Evidence · Hardware tests · Hypothesis · Branches

Directions analysis

SLAM — Simultaneous Localization and Mapping: estimate sensor motion while building a map from its observations.

• Robust localization and SLAM in dynamic environments: estimate motion and maintain a stable map despite moving objects.
• Semantic mapping, visual localization and navigation: describe objects and their locations, estimate pose from images, and reach a target.

Five elements organize this process:

Sensor data → Front end → Back end (optimize) → Map build
                  ↓              ↑
             Loop detection ─────┘

Element	Benefit	Dynamic SLAM	Semantic mapping
Sensor data	Cleaner input	Motion cues	RGB-D alignment
Front end	Stable matches	Static features	Object association
Back end	Consistent poses	Outlier rejection	Object alignment
Loop detection	Less drift	Place recognition	Relocalization
Map build	Stable maps	Dynamic removal	Object identities

The first direction emphasizes reliable motion and geometry. The second adds object meaning and target retrieval. The completed mapping experiments use supplied poses to isolate map construction. Complete SLAM and navigation remain outside these experiments. Both directions rely on multi-frame alignment: pose errors can corrupt static-structure filtering and object association, motivating the following open question.

open questions:

How can multi-frame mapping prevent localization errors from causing persistent mistakes in static-structure filtering and object association?

Related works

These are saved-result replays from supplied-pose mapping runs. GIFs preserve the companion MP4 timing; each selected map observation is held for eight frames at 12 fps. GIF timing and sources.

DUFOMap

KITTI-00 outdoor scene, 141 LiDAR scans. Observed empty space identifies dynamic points; pose margins protect static geometry. The GIF selects 21 scans to compare input, removed and retained points against the final map. MP4.

BeautyMap

KITTI-00 outdoor scene, 141 LiDAR scans. Occupancy comparisons remove dynamic traces; restoration protects static geometry. The GIF selects 21 scans to compare input, removed and retained points against the final map. MP4.

ConceptGraphs

Replica room0 indoor scene, 40 RGB-D frames. Geometry/CLIP matching fuses observations into 39 object representations. The GIF selects 20 frames, showing image segments, the final map and a red text-query candidate. Candidate correctness is unverified. MP4.

HOV-SG

In progress

Reproduction results

Work	Data	Result
DUFOMap	KITTI-00, 141 scans	Table IV accuracy matched
BeautyMap	Historical KITTI-02, 91 scans	Table III accuracy matched
ConceptGraphs	room0, 400 observations	Single-scene scored
HOV-SG		In progress

Evidence

Values below come from the papers linked by the original repositories and our recorded author-code runs. Each table concerns one method; no cross-method ranking is implied.

Shared metrics. DUFOMap and BeautyMap use the same SA/DA definitions. $S,D$ count ground-truth static/dynamic points; $S_k,D_r$ count correctly kept static and removed dynamic points. AA (DUFOMap) is the geometric mean; HA (BeautyMap) is the harmonic mean. DUFOMap definition · BeautyMap definition.

$$ \begin{aligned} \mathrm{SA} &= 100\frac{S_k}{S}, & \mathrm{DA} &= 100\frac{D_r}{D},\\ \mathrm{AA} &= \sqrt{\mathrm{SA},\mathrm{DA}}, & \mathrm{HA} &= \frac{2,\mathrm{SA},\mathrm{DA}}{\mathrm{SA}+\mathrm{DA}}. \end{aligned} $$

DUFOMap

KITTI-00, 141 released scans; full setting: voxel 0.1 m, d_s=0.2 m, d_p=1. Original repository · Paper Table IV · Our results.

Metric	Paper %	Reproduced %
SA	97.96	97.9635
DA	98.72	98.7196
AA	98.34	98.3408

All three full-setting values match the paper at two decimals. Across all five Table IV settings, all 15 accuracy entries match; runtime and online experiments are outside this comparison.

BeautyMap

Historical KITTI-02, frames 860–950, 91 scans; XY=1 m, Z=0.5 m, range=40 m. Original repository · Paper Table III · Our results.

Metric	Paper %	Reproduced %
SA	83.40	83.3978
DA	82.41	82.4092
HA	82.90	82.9006

All three values match at two decimals. All nine accuracy entries across XY=0.5/1/2 m match. This uses historical preprocessing/GT and the author's HA scorer; the exact paper method commit remains unidentified. Other sequences and runtime are outside this comparison.

ConceptGraphs

The paper reports Replica benchmark results; our completed result below covers room0 only, 400 observations, with the disclosed SAM batch-16 variant. These scopes differ, so the paper values are reference values, without a reproduction-gap calculation. Original repository · Paper Table II · Our room0 results.

Metric	Paper benchmark %	Our room0 %
mAcc	40.63	38.3156
F-mIoU	35.95	50.1379

In the author's evaluation confusion matrix, $n_c$ counts reference labels for class $c$, $p_c$ predicted labels and $t_c$ correct matches; $C$ is the evaluated class count and $N=\sum_c n_c$.

$$ \begin{aligned} \mathrm{IoU}_c &= \frac{t_c}{n_c+p_c-t_c}, & \mathrm{mAcc} &= \frac{100}{C}\sum_c\frac{t_c}{n_c},\\ \mathrm{mIoU} &= \frac{100}{C}\sum_c\mathrm{IoU}_c, & \text{F-mIoU} &= 100\sum_c\frac{n_c}{N}\mathrm{IoU}_c. \end{aligned} $$

Zero-denominator terms are 0. The author's mrecall is mAcc; F-mIoU is frequency-weighted, whereas our macro mIoU is 21.3460%. Scoring code.

Hardware tests

Desktop setup: RealSense D435 depth camera (without IMU) and Unitree L2 LiDAR.

Device	Basic test	Full video
D435	RGB / depth / stereo IR capture	24.6 s
L2	Point cloud / ICP display	42.2 s

Supplement: D435 RGB-D odometry, 44.9 s. These are recorded acquisition and display checks; hardware accuracy, full SLAM and cross-sensor fusion remain unverified. Timing and sources.

Hypothesis

We hypothesize that retaining the observation evidence behind map updates and revisiting these updates as pose estimates improve will reduce persistent mapping errors and preserve more consistent geometric and semantic maps

Branches

Branch	Role
main	Curated submission: question, four mapping reproductions, open question and hypothesis.
reproduce/author-originals	Original pipelines, paper-table/semantic scoring and recordings.
notes/personal-study-guide-20261008	Personal study notes and hardware trials under src/physical/, including operations and failures.

AI use · Sources/licenses · Citation · License