ProOOD: Lightweight 3D Occupancy Framework

• ProOOD is a lightweight plug-and-play framework that uses prototype-guided semantic imputation and tail mining to improve 3D semantic occupancy prediction under long-tailed data imbalance.
• It refines coarse 3D features by integrating global class prototypes and specialized modules (PGSI and PGTM) to boost tail-class representations and calibration.
• EchoOOD offers training-free, voxel-level OOD scoring by combining local logit coherence and prototype matching, ensuring reliable anomaly detection in autonomous driving.

ProOOD is a lightweight, plug-and-play framework for 3D semantic occupancy prediction under long-tailed category imbalance and out-of-distribution (OOD) inputs. It was introduced to address two coupled failure modes in autonomous driving: weak tail-class representations caused by real dataset imbalance, and voxel-level OOD absorption in which anomalies are overconfidently assigned to rare in-distribution classes. The method unifies prototype-guided refinement during training with training-free voxel-level OOD scoring at inference, through three components: prototype-guided semantic imputation (PGSI), prototype-guided tail mining (PGTM), and EchoOOD (Zhang et al., 1 Apr 2026).

1. Problem setting and motivation

3D semantic occupancy prediction produces voxel-wise semantic maps with geometry, and is central to autonomous driving. In the formulation used by ProOOD, the relevant setting is not limited to in-distribution urban scenes, but includes long-tailed class frequencies and unknown inputs. The paper identifies two interacting sources of unreliability (Zhang et al., 1 Apr 2026).

The first is long-tailed bias. Real datasets such as SemanticKITTI contain many rare classes, including truck at 0.16%, traffic-sign at 0.08%, and bicycle at 0.03%. Training on such imbalanced data yields weak tail-class representations and poor calibration. As a result, ambiguous voxels are overassigned to dominant classes, while rare categories remain poorly modeled.

The second is OOD absorption and overconfidence. Conventional post-hoc scores such as entropy, energy, and maximum softmax do not leverage 3D structure or semantic context, and therefore do not provide voxel-level uncertainty modeling tailored to volumetric prediction. Under distribution shift, anomalies such as construction barriers or clothing are often misclassified as rare in-distribution classes with high confidence.

ProOOD explicitly couples long-tail learning with voxel-level OOD detection. Its stated goals are to raise tail-class modeling capacity, improve calibration, reduce misattribution of anomalies to rare in-distribution classes, and provide reliable per-voxel OOD scores by quantifying local logit coherence and semantic feature consistency with class prototypes. This suggests that ProOOD treats representation quality and anomaly sensitivity as jointly constrained rather than separable subproblems.

2. Architecture and prototype formulation

ProOOD is integrated into a standard camera-based occupancy backbone such as SGN or VoxDet. The backbone extracts 2D multi-scale features, lifts them into coarse 3D features via view transformation, and then refines them with ProOOD modules before the original occupancy head performs semantics and occupancy prediction. The paper specifies the following flow (Zhang et al., 1 Apr 2026):

1. Build and update global class prototypes online with exponential moving average (EMA) from refined features of ground-truth voxels.
2. Apply PGSI to refine coarse features by imputation guided by global prototypes in occluded regions.
3. Apply PGTM to mine tail-class candidates at the coarse stage and strengthen them through prototype aggregation and dedicated supervision at the refined stage.
4. Feed refined features to the original occupancy head.
5. At inference, compute EchoOOD scores through local and global prototype matching together with local logit coherence.

The representation uses a 3D voxel grid of size $(H, W, Z)$, a feature volume $F \in \mathbb{R}^{H\times W\times Z\times C}$, logits $$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$, predicted labels $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$, and one global EMA prototype per non-empty semantic class, $p_k^{g} \in \mathbb{R}^{C}$.

Global prototypes are updated online using refined features from ground-truth non-empty voxels. With $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$ and refined feature $x_i^{(t)}$, ProOOD defines the batch mean and EMA update as

$$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$

Here $\operatorname{sg}(\cdot)$ denotes stop-gradient, and empty class $0$ is excluded. Wherever cosine similarity is used, prototypes are L2-normalized.

Prototype usage is controlled by a maturity gate based on intra-class variance:

$F \in \mathbb{R}^{H\times W\times Z\times C}$0

A prototype is consulted only if $F \in \mathbb{R}^{H\times W\times Z\times C}$1, $F \in \mathbb{R}^{H\times W\times Z\times C}$2, and $F \in \mathbb{R}^{H\times W\times Z\times C}$3, with $F \in \mathbb{R}^{H\times W\times Z\times C}$4 once past warm-up. This maturity mechanism is central because PGSI, PGTM, and EchoOOD all depend on prototype reliability.

3. Prototype-guided refinement during training

PGSI: prototype-guided semantic imputation

PGSI targets occluded or unobserved voxels that have non-trivial occupancy likelihood but are not visible in projections. The unobserved set is defined as

$F \in \mathbb{R}^{H\times W\times Z\times C}$5

where $F \in \mathbb{R}^{H\times W\times Z\times C}$6 is an auxiliary occupancy head and $F \in \mathbb{R}^{H\times W\times Z\times C}$7 denotes visible voxels from a depth-guided proposal (Zhang et al., 1 Apr 2026).

For each voxel in $F \in \mathbb{R}^{H\times W\times Z\times C}$8, PGSI computes prototype attention using Euclidean distance with temperature $F \in \mathbb{R}^{H\times W\times Z\times C}$9:

$$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$0

The feature is then updated by residual imputation,

$$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$1

The stated role of PGSI is to align filled content with class-level semantics, complementing geometry-based completion. A plausible implication is that PGSI treats occlusion recovery as a semantic estimation problem rather than only a geometric hallucination problem.

PGTM: prototype-guided tail mining

PGTM strengthens rare-class representations by mining candidate tail voxels at the coarse stage and reinforcing them via prototype aggregation. Using PGSI-updated coarse features, ProOOD computes normalized features and cosine similarity:

$$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$2

Let $$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$3 be the top-2 cosine margin. Tail candidates are filtered over tail classes $$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$4 by similarity $$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$5, margin $$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$6, and then TopK selection:

$$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$7

For each selected voxel, PGTM aggregates tail prototypes with temperature $$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$8:

$$L \in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}$$9

and updates the feature through a lightweight MLP $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$0:

$$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$1

At the refined stage, mined voxels are supervised by a dedicated head $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$2 with cross-entropy:

$$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$3

PBCL: prototype-based contrastive learning

ProOOD also introduces a prototype-aligned contrastive loss on refined non-empty voxels $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$4:

$$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$5

The framework keeps the original backbone’s occupancy and semantic losses unchanged and adds $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$6 and $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$7. The paper states that these regularizers improve calibration by reducing overconfidence on tail categories, which is reflected in ECE reductions.

4. EchoOOD and voxel-level OOD scoring

EchoOOD is the inference-time OOD component of ProOOD. It is training-free and parameter-free, and operates on refined features $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$8 and logits $$\hat{Y} \in \{0,\dots,K_{\mathrm{cls}}\}^{H\times W\times Z}$$9 over $p_k^{g} \in \mathbb{R}^{C}$0 classes. Only non-empty voxels are scored; empty class voxels receive the minimum fused score in the scene (Zhang et al., 1 Apr 2026).

The first cue is local logit coherence. For each predicted class $p_k^{g} \in \mathbb{R}^{C}$1, the mean logit vector is

$p_k^{g} \in \mathbb{R}^{C}$2

where $p_k^{g} \in \mathbb{R}^{C}$3 is the set of voxels predicted as class $p_k^{g} \in \mathbb{R}^{C}$4. The voxel score is

$p_k^{g} \in \mathbb{R}^{C}$5

The second cue is local prototype matching. Local scene prototypes are built from confident voxels per class:

$p_k^{g} \in \mathbb{R}^{C}$6

where $p_k^{g} \in \mathbb{R}^{C}$7 is the softmax top-2 probability gap. The corresponding score is

$p_k^{g} \in \mathbb{R}^{C}$8

The third cue is global prototype matching:

$p_k^{g} \in \mathbb{R}^{C}$9

Each component is min–max normalized over the scene, and fusion uses maximum aggregation:

$$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$0

For evaluation, thresholds are swept to compute AuROC and AuPRC$$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$1, so no fixed threshold is required. The design rationale stated in the paper is that maximum fusion is training-free and robust to long-tail bias. This suggests that EchoOOD is intended to flag anomalies whenever any one of the three signals becomes strongly inconsistent, rather than requiring consensus across cues.

5. Implementation, datasets, and evaluation protocol

ProOOD is inserted as a refinement block between coarse 3D features and the backbone’s 3D encoder. The paper reports compatibility with SGN, VoxDet, CGFormer, and VoxFormer. For SGN, the 2D encoder is ResNet-50 with FLoSP lifting. For VoxDet and CGFormer, the 2D encoder is EfficientNet-B7 or ResNet-50, with lifting via LSS depth and deformable attention. For CGFormer and VoxDet, voxel queries use resolution $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$2 with $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$3 channels (Zhang et al., 1 Apr 2026).

Default hyperparameters reported in the experiments include EMA momentum $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$4; warm-up iterations $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$5 for SGN and $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$6 for CGFormer/VoxDet on SemanticKITTI and KITTI-360 respectively; prototype quality thresholds $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$7, $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$8, and $$\Omega_k^{\mathrm{gt},(t)} = \{ i \mid y_i = k \text{ in current batch} \}$$9; PGSI initialization $x_i^{(t)}$0 with auxiliary occupancy loss weight $x_i^{(t)}$1; PGTM thresholds $x_i^{(t)}$2 and $x_i^{(t)}$3; and $x_i^{(t)}$4 ratios of $x_i^{(t)}$5 for SemanticKITTI and $x_i^{(t)}$6 for KITTI-360. EchoOOD uses $x_i^{(t)}$7 for local prototype confidence and max fusion without learned coefficients.

Training follows the backbone setup with AdamW and cosine annealing. The paper gives examples: SGN is trained on 4×RTX 3090 with batch size 4, AdamW with learning rate $x_i^{(t)}$8 and weight decay $x_i^{(t)}$9; CGFormer and VoxDet use cosine schedule, learning rate $$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$0, weight decay $$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$1, $$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$2, and $$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$3.

The evaluation protocol spans both in-distribution occupancy prediction and OOD detection.

Setting	Datasets	Metrics
3D occupancy prediction (ID)	SemanticKITTI; SSCBench-KITTI-360	IoU, mIoU, tail mIoU
OOD detection (OccOoD)	VAA-KITTI; VAA-KITTI-360; VAA-STU	AuPRC$$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$4, AuROC

SemanticKITTI contains 20 classes, specified as 19 objects plus free, with 10 train, 1 val, and 11 test sequences. SSCBench-KITTI-360 contains 19 classes, specified as 18 objects plus free, with 7 train, 1 val, and 1 test sequences. Tail mIoU is computed over classes below frequency thresholds, for example below 0.60% in SemanticKITTI and below 0.30% in KITTI-360. For OOD detection, VAA-KITTI and VAA-KITTI-360 are derived from SemanticKITTI and KITTI-360 with synthetic anomalies and contain 500 test images each in the single-frame setting; VAA-STU contains genuine anomalies from STU and has 578 test images. AuPRC$$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$5 uses ground-truth dilation radii $$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$6.

6. Empirical results, ablations, and efficiency

The paper reports state-of-the-art performance on both in-distribution occupancy prediction and OOD detection (Zhang et al., 1 Apr 2026). On SemanticKITTI, the headline result is that ProOOD surpasses baselines by +3.57% overall mIoU and +24.80% tail-class mIoU. On VAA-KITTI, it improves AuPRC$$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$7 by +19.34 points.

For in-distribution occupancy prediction, the reported examples include the following.

Dataset / Backbone	Baseline	ProOOD
SemanticKITTI / VoxDet mIoU	17.77	18.12
SemanticKITTI / VoxDet tail mIoU	6.17	6.34
KITTI-360 / SGN mIoU	46.22	46.62
KITTI-360 / SGN tail mIoU	8.17	8.35
KITTI-360 / VoxDet mIoU	48.22	48.23
KITTI-360 / VoxDet tail mIoU	10.92	11.52

For OOD detection with EchoOOD, the reported VAA-KITTI single-frame results on an SGN backbone are AuPRC$$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$8 and AuROC $$\begin{aligned} \mathbf{m}_k^{(t)} &= \frac{1}{|\Omega_k^{\mathrm{gt},(t)}|} \sum_{i\in\Omega_k^{\mathrm{gt},(t)}} \operatorname{sg}\big(\mathbf{x}_i^{(t)}\big),\ \gamma_k^{(t)} &= \beta \cdot \mathbf{1}\{|\Omega_k^{\mathrm{gt},(t)}| > 0\},\ \mathbf{p}_k^{g,(t)} &= \mathbf{p}_k^{g,(t-1)} + \gamma_k^{(t)}\big(\mathbf{m}_k^{(t)} - \mathbf{p}_k^{g,(t-1)}\big). \end{aligned}$$9. The comparison to ASS on OccOoD is $\operatorname{sg}(\cdot)$0 and AuROC $\operatorname{sg}(\cdot)$1. On VAA-STU with ProOOD plus SGN, the paper reports AuPRC$\operatorname{sg}(\cdot)$2 and AuROC $\operatorname{sg}(\cdot)$3.

The ablation study attributes separate effects to the major components. On SemanticKITTI validation, baseline to +PBCL yields tail mIoU +9.09% relative. Adding PGSI improves IoU and mIoU through better occlusion completion. Adding PGTM yields mIoU +1.42% and tail mIoU +6.77% relative, confirming effective tail mining. The anomaly-score comparison on VAA-KITTI shows EchoOOD outperforming post-proc, energy, entropy, and ASS by more than 12% AuPRC$\operatorname{sg}(\cdot)$4 across radii.

Sensitivity analyses also identify stable defaults. EMA momentum $\operatorname{sg}(\cdot)$5 is reported to balance stability and responsiveness, reaching mIoU 14.97 and tail mIoU 4.89. The default tail-mining thresholds $\operatorname{sg}(\cdot)$6 and $\operatorname{sg}(\cdot)$7 yield the best tail mIoU trade-off; too small $\operatorname{sg}(\cdot)$8 harms separation, whereas too large $\operatorname{sg}(\cdot)$9 excludes informative tail voxels. Prototype warm-up at $0$0 gives the best tail mIoU, avoiding both poor early initialization and delayed prototype benefits.

Calibration results are reported on VAA-KITTI with the SGN backbone: $0$1 decreases from 51.10 to 47.28, $0$2 from 57.49 to 53.91, and $0$3 from 4.93 to 4.41, while AuROC increases from 60.51 to 64.31. The paper interprets this as reduced overconfidence, especially on tail categories.

The efficiency overhead is limited. For SGN, ProOOD adds +0.28M parameters, changes FLOPs from 522.34G to 580.66G, changes FPS from 7.6 to 7.0, and changes inference memory from 5.16GB to 5.30GB. For VoxDet, it adds +0.01M parameters, +0.40G FLOPs, changes FPS from 5.0 to 4.9, and changes memory from 4.34GB to 4.47GB. EchoOOD itself is training-free and parameter-free.

7. Limitations, integration guidance, and broader significance

The paper identifies several limitations. Small or distant OOD objects may be missed if occupancy activation fails, because EchoOOD relies on occupancy features. Heavy occlusion and ambiguous 2D-to-3D lifting can cause false OOD responses in background regions. The framework also relies strongly on depth estimation quality, typically from an external mono or stereo module, so depth errors propagate into 3D features and prototypes. In addition, prototypes may inherit dataset biases, and extreme domain shifts or anomalies far outside learned semantics can reduce prototype utility (Zhang et al., 1 Apr 2026).

Several future directions are proposed: stronger or higher-resolution 2D backbones and object-aware refinement for small instances, end-to-end joint training with depth to reduce error propagation, losses that emphasize small-object voxels, adaptive prototype selection, multiple prototypes per class for intra-class diversity, weighted fusion in EchoOOD, and neighborhood-aware coherence beyond max aggregation. These suggestions indicate that the current single-prototype-per-class formulation is effective but not exhaustive.

The integration guidance given by the paper is operationally simple. The backbone is kept unchanged; PGSI and PGTM are inserted after view transformation or coarse 3D features; a per-class global prototype memory is maintained with EMA during training from refined features of ground-truth non-empty voxels; prototype usage is gated until maturity; and $0$4 and $0$5 are added while standard semantic and occupancy losses remain unchanged. At inference, EchoOOD constructs local prototypes from confident predictions, computes local logit and prototype scores together with the global prototype score, applies min–max normalization and max fusion, and produces voxel-level anomaly maps evaluated by AuPRC$0$6 and AuROC.

Within 3D OccOoD, ProOOD is defined by its explicit coupling of prototype-guided semantic refinement and training-free anomaly scoring. Its main technical claim is that improving class-consistent occupancy representations, particularly for tail classes and occluded regions, also improves voxel-level OOD detection and calibration. This coupling is the central conceptual contribution of the method.