Hierarchical Local Semantic Exclusion (HLSE)

• HLSE is a semantic-structure-aware mechanism that partitions global label spaces into locally exclusive clusters for robust evidential learning in BEV segmentation.
• It replaces the global Dirichlet simplex assumption with a local exclusivity constraint, better accommodating overlapping semantic classes such as drivable areas and lane markings.
• Integrated within NRSeg's dual-branch framework, HLSE improves uncertainty estimation and segmentation performance, particularly under domain shifts.

Searching arXiv for the cited papers to ground the article in current records. Hierarchical Local Semantic Exclusion (HLSE) is a semantic-structure-aware training mechanism introduced in the context of Birds' Eye View (BEV) semantic segmentation to make evidential uncertainty modeling compatible with label spaces that are not globally mutually exclusive. In NRSeg, HLSE addresses the mismatch between a global Dirichlet simplex and BEV semantics by partitioning classes into semantically defined local exclusive clusters and applying evidential learning within each cluster rather than across all classes at once (Li et al., 5 Jul 2025). More broadly, the term has also become useful as an interpretive lens for adjacent methods that isolate, suppress, or avoid semantically incompatible local structures within a hierarchy, even when those methods do not formalize HLSE by name (Qiu et al., 28 Dec 2025).

1. Conceptual definition

HLSE is motivated by a specific failure mode of structured prediction: some semantic domains are globally non-exclusive even though they contain locally exclusive subsets. In BEV semantic segmentation, this arises because the per-cell semantic map is represented as multiple channels, yet the semantics do not obey a single global categorical constraint. NRSeg makes this point explicitly: classes such as drivable area and lane markings can coexist spatially, and related overlaps may occur among drivable area, stop line, divider, pedestrian crossing, walkway, and carpark area under the adopted rasterization conventions (Li et al., 5 Jul 2025).

Within that setting, HLSE replaces a single global exclusivity assumption with a local one. Its formal definition is:

$$G =\{G_m,...,G_z\},\quad G_m=\{C_0,...,C_n\},\ (C_0 \cap \dots \cap C_n = 0).$$

Here, $G$ is the set of semantic exclusive local clusters, $G_m$ is one cluster, and $C_0,\dots,C_n$ are classes in that cluster. The condition $(C_0 \cap \dots \cap C_n = 0)$ means that classes inside a given cluster are mutually exclusive, while the full label space need not be globally exclusive (Li et al., 5 Jul 2025).

The term “local” in HLSE refers to locality in semantic space, not to image coordinates, spatial windows, adjacency graphs, or patch neighborhoods. The term “hierarchical” refers to the paper’s claim that semantic classes are “hierarchically grouped based on their inherent properties, forming multi-level local clusters,” although the exact hierarchy structure is not specified in the published description (Li et al., 5 Jul 2025). This suggests that HLSE is best understood not as a spatial operator, but as a structured decomposition of the label simplex into smaller semantically valid competitions.

2. Problem setting in NRSeg

NRSeg studies BEV semantic segmentation for autonomous driving using multi-view camera images, perspective-view encoding, BEV transformation, temporal fusion, and BEV prediction heads. The implementation details specify 6 static classes on nuScenes: drivable area, pedestrian crossing, walkway, stop line, carpark area, and divider. The BEV space covers $(-50,50)$ meters with resolution $0.5$ m, and the base architecture is described as an LSS-style BEV segmentation pipeline with perspective encoder, view transformer, BEV decoder or segmentation heads, and streaming temporal fusion (Li et al., 5 Jul 2025).

The paper introduces HLSE inside a larger framework, NRSeg, which has two major components. The first is PGCM, a Perspective-Geometry Consistency Metric used to quantify the guidance capability of synthetic generated data and to reweight segmentation supervision. The second is BiDPP, a Bi-Distribution Parallel Prediction framework that learns both a multinomial-like semantic probability branch and a Dirichlet evidential branch (Li et al., 5 Jul 2025).

HLSE belongs to the Dirichlet half of BiDPP. This placement is essential. Evidential deep learning assumes mutually exclusive categories because the Dirichlet distribution is defined over a simplex. NRSeg explicitly shows why that assumption fails for globally overlapping BEV semantics. If two non-exclusive classes are modeled in a single Dirichlet, their expected probabilities are

$$P_e^{1} = \alpha_{1}/(\alpha_{1}+\alpha_{2}), \qquad P_e^{2} = \alpha_{2}/(\alpha_{1}+\alpha_{2}).$$

If both classes should be active, neither expected value can reach $1$, so the prior is structurally inappropriate (Li et al., 5 Jul 2025). HLSE is therefore the mechanism that makes the evidential branch semantically valid in this domain.

A common misconception is to treat HLSE as a feature-extraction block inserted into the encoder or decoder. NRSeg does not present it that way. HLSE is described as a training module or structural loss design acting on the Dirichlet prediction head outputs. It does not operate on feature-map neighborhoods, local windows, post-processed masks, or logits from the multinomial branch (Li et al., 5 Jul 2025).

3. Formal mechanism and losses

In NRSeg, the shared BEV network outputs two prediction heads,

$D_m^k, D_d^k = N(F_b^{t'}),$

where $G$0 is the multinomial branch output and $G$1 is the Dirichlet or evidential branch output (Li et al., 5 Jul 2025). HLSE acts on the Dirichlet branch. Evidence and Dirichlet parameters are defined as

$G$2

The Dirichlet density is written as

$G$3

with expected class probability and uncertainty

$G$4

HLSE then restricts evidential learning to each local cluster $G$5. For a given cluster, the paper states that parameterization is implemented by taking the corresponding subset of $G$6, written as $G$7, and applying the evidential loss only within that subset (Li et al., 5 Jul 2025).

The cluster-level evidential loss is

$G$8

where $G$9 indexes BEV pixels, $G_m$0 indexes classes within cluster $G_m$1, and $G_m$2 is the sum of the cluster’s Dirichlet parameters. A KL regularizer is added:

$G_m$3

The total HLSE evidential objective is

$G_m$4

This loss is then inserted into the source and target training objectives through $G_m$5 and $G_m$6 (Li et al., 5 Jul 2025).

Algorithmically, the paper describes the following steps: convert $G_m$7 to $G_m$8; for each cluster $G_m$9, gather the relevant class channels and labels; compute $C_0,\dots,C_n$0; compute $C_0,\dots,C_n$1; evaluate the evidential and KL losses; and accumulate them into $C_0,\dots,C_n$2 (Li et al., 5 Jul 2025). The exact class-group assignments, however, are not specified.

4. Role within the full NRSeg objective

HLSE does not replace the ordinary segmentation branch. NRSeg uses a multinomial branch supervised by a DICE-style loss and an evidential branch supervised by HLSE. The multinomial branch predicts

$C_0,\dots,C_n$3

Its BEV loss is

$C_0,\dots,C_n$4

and for synthetic data this denominator is modified by the PGCM-derived factor $C_0,\dots,C_n$5 (Li et al., 5 Jul 2025).

The full source and target objectives are

$C_0,\dots,C_n$6

$C_0,\dots,C_n$7

and

$C_0,\dots,C_n$8

The reported hyperparameters are $C_0,\dots,C_n$9, $(C_0 \cap \dots \cap C_n = 0)$0, $(C_0 \cap \dots \cap C_n = 0)$1, $(C_0 \cap \dots \cap C_n = 0)$2 as a sigmoid ramp-up from $(C_0 \cap \dots \cap C_n = 0)$3 to $(C_0 \cap \dots \cap C_n = 0)$4, and $(C_0 \cap \dots \cap C_n = 0)$5 for the KL term (Li et al., 5 Jul 2025).

This design clarifies the division of labor. PGCM addresses noisy synthetic-label alignment. HLSE addresses invalid exclusivity assumptions in evidential uncertainty learning. BiDPP combines both by pairing a flexible multi-label segmentation branch with a clusterwise Dirichlet uncertainty branch (Li et al., 5 Jul 2025).

At inference time, HLSE remains primarily a training-time mechanism, but the learned Dirichlet branch is used to derive uncertainty and filtered predictions. The paper states that “Ours-D denotes the low-uncertainty outcomes from the Dirichlet distribution” and reports thresholding uncertainty below $(C_0 \cap \dots \cap C_n = 0)$6 for visualization (Li et al., 5 Jul 2025).

5. Empirical evidence, scope, and limitations

The strongest direct quantitative evidence for HLSE is indirect, because NRSeg does not provide an ablation isolating HLSE from BiDPP. The paper explicitly notes that the ablation tables report BiDPP as a whole rather than “BiDPP without HLSE” versus “BiDPP with HLSE” (Li et al., 5 Jul 2025).

The closest available numbers come from the core-module ablations. In unsupervised domain adaptation from Boston to Singapore, adding BiDPP on top of Stream + synthetic data + PGCM improves mIoU from $(C_0 \cap \dots \cap C_n = 0)$7 to $(C_0 \cap \dots \cap C_n = 0)$8, a gain of $(C_0 \cap \dots \cap C_n = 0)$9. In Day to Night adaptation, the improvement is from $(-50,50)$0 to $(-50,50)$1, again $(-50,50)$2. In semi-supervised learning with $(-50,50)$3 labels, adding BiDPP improves mIoU from $(-50,50)$4 to $(-50,50)$5, a gain of $(-50,50)$6. The paper summarizes PGCM and BiDPP as delivering gains of $(-50,50)$7 and $(-50,50)$8, respectively, in Boston $(-50,50)$9 Singapore (Li et al., 5 Jul 2025).

The evidence therefore supports a modest but consistent contribution from the Dirichlet branch whose validity depends on HLSE. The visual analysis further suggests that the evidential branch is helpful for detailed predictions and for avoiding uncertain mistakes under domain shift. One example described in the paper highlights a nighttime region where high uncertainty helps avoid classifying an area as walkway (Li et al., 5 Jul 2025).

Several implementation details remain unspecified. The paper does not provide the actual cluster assignments for the 6 nuScenes classes, does not show a tree or parent-child hierarchy for the semantic groups, does not state whether clusters overlap, and does not fully define $0.5$0 in the KL term. It also does not report FLOPs, runtime, or memory overhead for HLSE separately (Li et al., 5 Jul 2025). These omissions are material for exact reproduction.

A second common misconception is to read “local” as spatial locality. The paper provides no local patch radius, no adjacency graph over pixels, and no exclusion mask over feature neighborhoods. The locality is purely semantic: local class clusters that are mutually exclusive within the cluster but not necessarily outside it (Li et al., 5 Jul 2025).

6. Relations to adjacent hierarchical exclusion ideas

Although HLSE is explicitly defined in NRSeg, several other recent arXiv papers implement closely related mechanisms under different names. These are best treated as analogues rather than identical formulations.

In HeroSQL, semantic validation for Text-to-SQL is cast as a hierarchical representation problem combining Logical Plans for global intent and Abstract Syntax Trees for local detail. The method does not define a local semantic exclusion operator, but it does localize semantically incorrect sub-SQL regions, map AST perturbations back to corresponding LP fragments, and use node-level logical-plan embeddings to identify anomalous substructures (Qiu et al., 28 Dec 2025). This suggests an HLSE-like interpretation in which local structures can be surfaced and treated as semantically untrustworthy without discarding the full-query context.

HCSC provides a different analogue in self-supervised vision. Its hierarchical prototypes and selective coding mechanism probabilistically filter semantically ambiguous negatives at multiple hierarchy levels. Instance negatives aligned with the query’s prototype are less likely to be kept, and prototype negatives aligned with the same parent are also less likely to be used (Guo et al., 2022). This is not HLSE by name, but it is a direct form of hierarchical semantic exclusion in the negative-sampling sense.

HiLight, in hierarchical text classification, avoids a structure encoder and instead uses the hierarchy to define local hard negatives—specifically siblings and descendants—for each positive label. Its local contrastive loss makes the positive label compete only with these nearby alternatives (Chen et al., 2024). Here again, the mechanism is exclusion-like because it strengthens discrimination among hierarchy-local confusions rather than across arbitrary global negatives.

SLoD extends the family of related ideas in a different direction. It does not define hard exclusion, but its heat-kernel diffusion on hyperbolic manifolds progressively attenuates fine local semantic detail as the scale parameter $0.5$1 increases, while preserving it as $0.5$2. The paper explicitly frames this as continuous resolution control and automatic boundary discovery across abstraction levels (Izgorodin, 9 Mar 2026). A plausible implication is that SLoD supplies a continuous, geometry-aware notion of semantic attenuation, whereas HLSE supplies a discrete structured-loss mechanism for local exclusivity.

Taken together, these works indicate that HLSE names one member of a broader methodological family: hierarchical methods that replace global semantic competition with locally valid exclusions. In NRSeg this is done by decomposing the Dirichlet simplex into semantic clusters; in HeroSQL by isolating anomalous sub-SQL regions; in HCSC by filtering false negatives hierarchically; in HiLight by local contrastive discrimination; and in SLoD by scale-dependent attenuation rather than hard masking (Li et al., 5 Jul 2025).