Height Interval Filtering (HIF)

• Height Interval Filtering (HIF) is a technique that partitions vertical domains into discrete intervals to filter, aggregate, and analyze data.
• It boosts computational efficiency and accuracy by restricting processing to relevant height intervals in robotics, 3D mapping, and vision systems.
• HIF leverages methods like Bayesian updates and data-driven partitioning to enhance dynamic point cloud mapping, occupancy prediction, and aerial visual place recognition.

Height Interval Filtering (HIF) is a class of techniques for partitioning the vertical (height) dimension of a domain—whether 3D spatial data, visual data, or stochastic fields—into discrete intervals to filter, aggregate, or analyze information with respect to height. HIF methods have been developed independently in multiple fields, including robotics, computer vision, and stochastic process theory, where they enable efficient computation, improve robustness to scale and occlusion, and sharpen measurement or retrieval tasks by minimizing irrelevant cross-height interference.

1. Formal Definition and General Principle

Height Interval Filtering refers to the process of dividing a (potentially continuous) height or vertical measurement domain $H$ into $L$ disjoint intervals $\left\{[h_{\ell-1}, h_\ell)\right\}_{\ell=1}^L$, and applying filtering, selection, or aggregation such that, for each query (e.g., sensor measurement, image, or spatial point), only data or operations relevant to the height interval(s) containing or neighboring the query's estimated height are considered. Formally, the partition is:

$H = \bigcup_{\ell=1}^{L} [h_{\ell-1}, h_\ell),$

where $0 = h_0 < h_1 < \dots < h_L = h_\text{max}$ and $\Delta h = h_\ell - h_{\ell-1}$ may be uniform or data-driven.

The core innovations of HIF lie in:

• reducing computational and memory burdens by restricting queries to selected height intervals,
• improving accuracy by filtering out data from incompatible height regimes,
• supporting online updating or learning via interval-wise probability or feature aggregation.

2. Applications and Implementations Across Domains

HIF manifests in different algorithmic forms tailored to application domains:

Robotics and 3D Mapping: Dynamic Points Removal

In dynamic point cloud mapping, HIF structures the environment as a 2D grid of vertical pillars, where each pillar stores a set of height intervals with associated probabilities representing static occupancy. Incoming LiDAR points are bucketed into pillars and assigned to height intervals. Bayesian updates are applied to interval occupancy probabilities based on the presence or absence of returns within those intervals. This structure enables efficient dynamic point removal, real-time updating, and robust handling of occlusion via a "low-height preservation" heuristic, which avoids erroneously marking unseen lower intervals as free when occlusions are suspected (2503.06863).

Vision-based 3D Occupancy Prediction

In monocular depth or occupancy networks, HIF is implemented by first predicting per-pixel height (using supervised height maps), followed by Mask-Guided Height Sampling (MGHS), which generates binary masks for predefined height intervals. 2D image features are filtered by these masks, then projected into corresponding height subspaces in the 3D grid, where Synergistic Feature Aggregation further refines features by leveraging both depth-based and height-stratified cues. Data-driven interval selection (e.g., via minimizing weighted average entropy over semantic occupancy classes) ensures that interval boundaries maximize semantic purity or relevance (Wu et al., 2024).

Aerial Visual Place Recognition (VPR)

In aerial navigation, HIF allows a query image to be matched only against map tiles captured within relevant flight-height intervals, reducing the impact of severe scale variance in appearance and drastically shrinking search space. A lightweight height-estimation adapter computes a height descriptor for the query, which is then matched against a small database of height-interval prototypes; only the top-$k$ height intervals are selected for downstream place-recognition. Subtle scale offsets are further mitigated by center-weighted masking of features. This decoupled structure also enables substantial memory and compute savings (e.g., up to 90% reduction in visual place recognition databases) with maintained or improved recall (He et al., 4 Mar 2026).

Stochastic Field Theory

In the probabilistic analysis of planar Gaussian fields, HIF describes the calculation of statistical quantities (e.g., the count of critical points) restricted to points whose function values fall within a given height interval. Here, HIF enables the derivation of sharp moment bounds for these height-restricted counts, including in the asymptotic regime where the interval shrinks with the domain size. The expected number of critical points within a shrinking window is precisely controlled via Kac–Rice formulae, with moment bounds depending critically on the width of the height interval (Muirhead, 2019).

3. Algorithmic Structures and Methodologies

The operational structure of HIF depends on the modality but shares common principles:

Height Interval Partition Definition

In all implementations, the domain is discretized along the vertical dimension. Intervals can be uniform (e.g., every 50 m for aerial VPR or grid-based mapping), statistically informed (e.g., entropy-minimizing partitions in occupancy prediction), or variable depending on analytic necessity (as in fields theory).

Filtering and Querying

For each query (LiDAR scan, image, etc.), an explicit or learned estimator predicts the relevant height interval(s). Only data or operations in those intervals are considered:

• Point cloud mapping: Bayes filter/occupancy update is run per height interval per pillar.
• 3D occupancy: Feature masking and projection occur for each interval, preventing cross-height confounding.
• VPR: Database search is limited to top-$k$ height intervals determined by learned height descriptors.

Probability and Feature Updates

Probabilistic filtering per interval is often employed. For example, in dynamic point removal, each interval in a pillar maintains a "static-occupancy" probability, updated via Bayesian inference after each scan, with special consideration for occlusion at low heights. In occupancy networks, feature aggregation uses affinities learned over channel and spatial axes, leveraging height-aware decomposition to suppress irrelevant features.

Complexity and Efficiency

HIF reduces complexity compared to non-partitioned approaches. In point cloud mapping, the pillar-plus-height-interval structure, combined with hash-based indexing, supports updates and filtering at approximately O($N + M \log M$) per scan ($N$ points, $M$ pillars), yielding a 7.7× speedup over voxel-based methods (2503.06863). Aerial VPR achieves up to a 90% reduction in memory usage via sub-database filtering (He et al., 4 Mar 2026).

4. Mathematical Formulations and Statistical Properties

Gaussian Field Example

Let $f:\mathbb{R}^2 \to \mathbb{R}$, a planar Gaussian field. Define the height-interval filtered critical point count:

$$N_D(I) = \#\{x\in D: \nabla f(x)=0\text{ and } f(x)\in I\}$$

where $D$ is a domain, $I$ is the interval. The mean count is given (under non-degeneracy) by a Kac–Rice integral:

$\mathbb{E}[N_D(I)] = |D|\int_I p(s)\,ds$

where $p(s)$ is the density for critical points at height $s$.

A second-moment bound (with $B_R$ a ball of radius $R$, $I_R$ an interval of width $\Delta_R$):

$\E[N_{B_R}(I_R)^2] \le c\min\{ R^4\Delta_R^2 + R^2\Delta_R,\ R^4\}$

Sharpness in the shrinking-window regime is obtained: for $\Delta_R\gg R^{-2}$, $\E[N^2]\asymp R^4\Delta_R^2$; if $\Delta_R\ll R^{-2}$, $\E[N^2]\lesssim R^2\Delta_R\ll1$ (Muirhead, 2019).

Dynamic Point Cloud Mapping

Each pillar $P_{mn}$ stores intervals $H_{mn} = \{(b_k, t_k, p_k)\}$ with probability $p_k$ updated as:

$$p_k^\text{new} = \frac{P_s p_k}{P_s p_k + P_p (1 - p_k)}$$

with $P_s, P_p$ set depending on presence/absence of observation ("BayesFilter" update) (2503.06863).

Vision-Based Occupancy

Given a height map $H_\text{pred}\in\mathbb{R}^{B\times h\times w}$ discretized into $N_h$ intervals, interval masks $H_\text{mask}^k$ and masked features $F_\text{ha}^k$ are obtained as:

$$H_\text{mask}^k(u,v) = \begin{cases} 1, & H_\text{map}(u,v) \in I_k \ 0, & \text{otherwise}\end{cases}$$

$$F_\text{ha}^k(u,v,c) = H_\text{mask}^k(u,v)\cdot F_\text{ctx}(u,v,c)$$

Projection and aggregation proceed per-interval, then outputs are fused via attention-like mechanisms (Wu et al., 2024).

Visual Place Recognition

The height-estimation branch produces a descriptor $g_\text{he}(q)$ for a query $q$, which is scored against height database entries via cosine similarity. The top-$k$ intervals drive selection of map sub-databases for subsequent VPR (He et al., 4 Mar 2026).

5. Empirical Impact and Benchmarks

Height Interval Filtering delivers substantial gains in efficiency and/or accuracy across domains:

• Dynamic Points Removal: HIF achieves 7–8× faster processing than voxel- or ray-casting-based baselines (80–90 FPS vs. 10–15 FPS), with Associated Accuracy (AA) within 1–2% of the best SOTA, as shown on KITTI and Argoverse datasets (2503.06863).
• Vision-Based Occupancy: The MGHS + SFA approach in DHD raises mIoU by +2.18 points when added to a ResNet50 baseline. Interval partitioning further boosts performance. The DHD-L variant achieves 45.53% mIoU compared to 43.52% for FlashOcc, demonstrating state-of-the-art results with modest extra complexity (Wu et al., 2024).
• Aerial VPR: HIF in HE-VPR reduces memory usage up to 90% and can surpass baseline Recall@1 (e.g., 70.42% vs. 69.50% on GEStudio) by restricting matching to the top-k relevant height intervals (He et al., 4 Mar 2026).
• Stochastic Fields: The second-moment bounds for the height-interval-critical-point count are sharp, allowing controlled analysis in shrinking or variable height windows (Muirhead, 2019).

6. Key Parameters and Implementation Considerations

HIF's effectiveness relies on principled selection of:

• Height interval width $\Delta h$ (domain- or data-adaptive),
• Number of intervals $L$ or $N_h$ (tradeoff between precision and computational cost),
• Bayesian parameters ($\alpha, \beta$) in dynamic mapping,
• Entropy or semantic criteria for interval partitioning in occupancy networks,
• Architectural choices (adapter hidden dimension, pooling mechanisms, feature fusion) in deep models,
• K value (number of intervals selected for downstream query, e.g., $k=5,10$ in HE-VPR).

Occultation heuristics such as low-height preservation increase robustness in dynamic/noisy environments. Hyperparameter values are empirically validated (e.g., $\Delta h=50$ m, $k=5$, image resolution $224\times224$ in HE-VPR), and modest-height databases (~100–850 entries) are sufficient in practical systems (He et al., 4 Mar 2026).

7. Connections, Generalizations, and Limitations

HIF conceptually generalizes to any domain where the vertical or scalar-valued dimension governs the relevance or semantic purity of observations. While originated in spatial intelligence and random field analysis, its principled structure enables integration into learned architectures, probabilistic filters, and analytic systems alike. Limitations arise with extremely sparse or ambiguous height measurements, or where the distinction between intervals is not preserved due to sensor error or environmental homogeneity. Nevertheless, across empirical and theoretical contexts, Height Interval Filtering represents a foundational methodology for scalable, robust, and precise vertical-dimension processing.