PAIRS Lab

Preprint 2026

Spatio-Temporal Semantic Mapping

Efficient continuous semantic mapping based on spatio-temporal awareness

My Le Pham1, Dinh Trieu Duong1, Xiem HoangVan1, Thanh Nguyen Canh1,2

1VNU University of Engineering and Technology, Hanoi   2Japan Advanced Institute of Science and Technology

arXiv Results
Semantic maps of SemanticKITTI sequence 05: camera image, CSM, the proposed method and ground truth, with zoomed regions where the proposed map is more coherent.
SemanticKITTI sequence 05: CSM, ours and ground truth. Ours has fewer scattered labels.

01 · Abstract

Infer where it is uncertain, remember for as long as it matters

Reliable continuous semantic mapping must integrate noisy LiDAR observations while keeping the map spatially accurate and temporally consistent. Existing methods often use a fixed spatial support, or accumulate semantic evidence without accounting for elapsed time, which wastes computation and leads to inconsistent labels.

We propose an uncertainty- and time-aware framework that combines adaptive spatial inference with temporal semantic fusion. Normalized Shannon entropy sets each voxel's kernel length: wider neighborhoods where the labels are uncertain, narrower ones where they are confident. An elapsed-time-dependent decay weights historical evidence across successive scans.

On SemanticKITTI sequences 03–06 the framework has the highest overall accuracy and weighted IoU of the evaluated methods. On sequence 03 it reaches 54.92% mIoU, 13.18 points above spatial-only inference, and the adaptive kernel is both more accurate and faster than every fixed kernel tested.

02 · Method

Adaptive spatial inference, then time-aware fusion

Pipeline: LiDAR points and labels are downsampled, voxelized and filtered; spatial inference estimates semantic uncertainty, picks an adaptive kernel length and runs kernel inference; temporal fusion decays the previous counter by elapsed time and fuses it into the continuous semantic map.
Fig. 1: overall architecture.
  1. 1

    Preprocessing

    LiDAR points with semantic labels are downsampled, voxelized in an octree, and redundant voxels are filtered out.

  2. 2

    Semantic uncertainty

    Normalized Shannon entropy H̄(v) ∈ [0, 1] of each voxel's class distribution, using all competing classes, not only the top one.

  3. 3

    Adaptive kernel length

    L(v) grows linearly with H̄(v): confident voxels use a small neighborhood, ambiguous ones gather more context.

  4. 4

    Kernel inference

    A compactly supported kernel accumulates per-class evidence from nearby points; points outside the support do not contribute.

  5. 5

    Time-aware fusion

    The per-voxel counter is updated as σt = α σt−1 + Δσt, with α = e−Δt/τ set by the time since the last update.

  6. 6

    Label

    Counters are normalized into class probabilities, and each voxel takes the most probable class.

Time-aware fusion: the previous per-class counter is decayed and added to the kernel-weighted evidence of the current scan to give the final class probabilities.
Fig. 3: time-aware fusion of the previous per-class counter with the current scan's kernel-weighted evidence.

03 · Temporal decay

How much does old evidence still count?

The decay factor α = e−Δt/τ (Eq. 6, Fig. 2) weights the previous counter. A larger time constant τ keeps old information longer; a smaller one forgets faster. Move the sliders.

The paper's Fig. 2 shows τ = 0.4, 1 and 3.

04 · Results

SemanticKITTI sequences 03–06

Compared with count-and-smooth mapping (CSM) and spatial Bayesian kernel inference (S-BKI). Pick a sequence and a metric or class.

The gains are not universal: on sequence 05, sidewalk IoU is lower than both baselines, as the paper notes.

05 · Adaptive kernel

More accurate and faster than any fixed kernel

Accuracy% · higher is better · Table 1
Runtime for 10 scanss · lower is faster · Table 1

Runtimes refer to the 10-scan benchmark of the paper and are not a claim of real-time performance.

06 · Ablation

What temporal fusion adds

07 · Qualitative

Semantic maps

Two rows of semantic maps on SemanticKITTI sequence 05 with zoomed insets: image, CSM, our method and ground truth.
Fig. 4: SemanticKITTI sequence 05. CSM (left), ours (center), ground truth (right). Ours has fewer isolated labels and more continuous structures.

Limitations stated in the paper: the method depends on the quality of the input segmentation and does not explicitly handle dynamic objects.

08 · Cite

Citation

@misc{pham2026spatiotemporal,
  title         = {Efficient Continuous Semantic Mapping based on Spatio-Temporal Awareness},
  author        = {Pham, My Le and Duong, Dinh Trieu and HoangVan, Xiem and Canh, Thanh Nguyen},
  eprint        = {2606.22672},
  archivePrefix = {arXiv},
  year          = {2026}
}