EchoOOD: Voxel-Level OOD Detector

• EchoOOD is a training-free mechanism that detects voxel-level anomalies by fusing local logit coherence, scene-specific prototypes, and global class prototypes.
• It operates at inference time within ProOOD to counter overconfidence in 3D semantic occupancy predictions under long-tailed class distributions.
• The method leverages cosine similarity and a max-aggregation rule to merge multiple discrepancy cues, yielding significant improvements on standard OOD benchmarks.

EchoOOD is a prototype-driven, training-free mechanism for voxel-level out-of-distribution detection in 3D semantic occupancy prediction, introduced as the OOD-scoring component of ProOOD. In that framework, a dense voxel grid is assigned semantic labels by an occupancy predictor, but anomalous voxels are often strongly and overconfidently assigned to known classes, especially under long-tailed class distributions. EchoOOD addresses this by scoring each non-empty voxel from three complementary perspectives—local logit coherence, local scene-specific prototype matching, and global class-prototype matching—and fusing those signals into a voxel-level anomaly map (Zhang et al., 1 Apr 2026).

1. Position within ProOOD and target failure mode

EchoOOD is defined inside ProOOD, whose three components are Prototype-Guided Semantic Imputation (PGSI), Prototype-Guided Tail Mining (PGTM), and EchoOOD. PGSI and PGTM are training-time refinement modules that improve the learned occupancy representation, especially for occluded regions and tail classes, while EchoOOD is applied at inference time on top of the resulting occupancy predictor. The paper explicitly characterizes EchoOOD as requiring “no additional parameters or training” and as an efficient plug-and-play solution (Zhang et al., 1 Apr 2026).

The motivating problem is overconfidence under long-tail bias in 3D semantic occupancy prediction. Current occupancy models are described as vulnerable to strongly and overconfidently assigning anomalous voxels to known classes, particularly rare classes with low feature density and high intra-class variance. Standard post-hoc scores such as maximum softmax probability, entropy, and energy are presented as limited in this setting because they are formulated for per-sample classification and do not account for the spatially structured, voxel-wise semantic context of 3D scenes. EchoOOD therefore operationalizes a different criterion: an in-distribution voxel should be self-consistent not only with its own classifier output, but also with class-consistent evidence in the current scene and with dataset-level class structure (Zhang et al., 1 Apr 2026).

In practical terms, EchoOOD is not a separate occupancy predictor. It consumes the refined features and semantic outputs of an occupancy model after PGSI and PGTM have shaped the representation space. This implies that its performance is coupled to prototype quality and class-feature geometry, even though the scoring rule itself is training-free at inference.

2. Input representation and three-branch scoring rule

At inference time, EchoOOD operates on a voxel grid of size $(H,W,Z)$. The occupancy model produces a refined voxel feature volume

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

and per-voxel semantic logits

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

For EchoOOD specifically, the feature map is the refined feature volume, denoted in the paper as $F^{3D}_{\text{refined}}$, together with semantic predictions $\hat{\mathbf{Y}}$ and the maintained global class prototypes $\mathbf{p}^{\mathrm g}_k$ (Zhang et al., 1 Apr 2026).

EchoOOD computes scores only for voxels with $\hat y_i>0$, that is, non-empty voxels. Its anomaly score is assembled from three branches.

Branch	Reference object	Core idea
Local logit score	Class-wise mean logit in the current scene	Penalizes logit patterns incoherent with the predicted class
Local prototype score	Scene-specific prototype from confident same-class voxels	Penalizes feature deviation from the current scene’s class cluster
Global prototype score	Global EMA class prototype	Penalizes feature deviation from the canonical class representation

The local logit score compares a voxel’s logit vector to the mean logit pattern of its predicted class in the current scene. For each class $k$, the class-wise mean logit is

$$\boldsymbol{\mu}_k = \frac{1}{|\Omega_k|}\sum_{j\in\Omega_k} \mathbf{l}_j.$$

For voxel $i$ with predicted class $$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$0, EchoOOD defines

$$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$1

This is one minus cosine similarity. A larger value indicates that the voxel’s raw logits are less aligned with the class-typical logit pattern in the current scene (Zhang et al., 1 Apr 2026).

The local prototype score is computed from scene-specific prototypes formed using confident same-class voxels. The local prototype for class $$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$2 is

$$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$3

with

$$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$4

The confidence filter uses the top-2 softmax probability gap

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

The score is then

$$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$6

This branch measures whether a voxel fits the scene-specific realization of its predicted class (Zhang et al., 1 Apr 2026).

The global prototype score compares the same voxel feature to the long-term class anchor:

$$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$7

Here the reference is not scene-local but accumulated across training scenes. The paper distinguishes these roles directly: global EMA prototypes are “stable class anchors across scenes,” while local prototypes “capture scene-local coherence” (Zhang et al., 1 Apr 2026).

3. Fusion, normalization, and empty-voxel handling

Each component score is min-max normalized to $$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$8 over the scene before fusion. EchoOOD then applies a maximum-aggregation strategy:

$$\mathbf{F}\in \mathbb{R}^{H\times W\times Z\times C}$$9

This is a hard OR over three discrepancy cues: if any branch indicates strong abnormality, the voxel receives a high fused anomaly score (Zhang et al., 1 Apr 2026).

The choice of maximum aggregation is methodologically important. The paper presents no learned fusion weights, additive combination, or temperature scaling inside EchoOOD itself. This makes the module parameter-free at inference and keeps it aligned with the broader ProOOD claim of a lightweight, plug-and-play post-hoc detector. A plausible implication is that the method is designed to favor sensitivity to any one semantic inconsistency rather than averaging away isolated but strong anomaly evidence.

A special rule is used for empty voxels, class $$\mathbf{L}\in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}.$$0. For voxels predicted as empty,

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

This conservative handling suppresses anomaly scores on empty voxels. The supplementary discussion notes that some road voxels can still be marked as OOD due to a “conservative strategy,” but the operational link between that qualitative remark and the empty-voxel rule is not fully elaborated in the provided text (Zhang et al., 1 Apr 2026).

4. Prototype construction and maturity gating

EchoOOD’s prototype machinery is inherited from ProOOD’s training modules rather than learned by a separate OOD head. The persistent prototypes are class-wise voxel prototypes in the learned feature space,

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

for each non-empty class $$\mathbf{L}\in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}.$$3. These are accumulated across training batches using only ground-truth voxels, excluding empty class $$\mathbf{L}\in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}.$$4, which the paper says ensures “semantic purity” (Zhang et al., 1 Apr 2026).

If $$\mathbf{L}\in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}.$$5 is the set of ground-truth voxels of class $$\mathbf{L}\in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}.$$6 in iteration $$\mathbf{L}\in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}.$$7, the EMA update is

$$\mathbf{L}\in \mathbb{R}^{H\times W\times Z\times K_{\mathrm{cls}}}.$$8

with

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

If no class-$F^{3D}_{\text{refined}}$0 voxels appear in the batch, the update is skipped. The paper also defines normalized prototypes for cosine-based operations:

$F^{3D}_{\text{refined}}$1

These are shared across PGSI, PGTM, PBCL, and EchoOOD rather than maintained as an EchoOOD-specific memory bank (Zhang et al., 1 Apr 2026).

The method additionally uses prototype maturity gating. Prototype quality is defined via intra-class variance:

$F^{3D}_{\text{refined}}$2

A class-$F^{3D}_{\text{refined}}$3 prototype is only consulted if training has progressed beyond a warm-up point and if quality and support-size conditions are met: $F^{3D}_{\text{refined}}$4, $F^{3D}_{\text{refined}}$5, and $F^{3D}_{\text{refined}}$6. The paper states that this gate applies to PGSI, PGTM, and to the global-prototype component of EchoOOD (Zhang et al., 1 Apr 2026).

This gating mechanism matters because EchoOOD does not blindly trust global prototypes. At the same time, the paper does not provide an explicit test-time fallback rule for classes whose global prototypes remain unavailable or immature. This suggests an implementation detail is intentionally left unspecified in the available description.

5. Integration, efficiency, and empirical behavior

EchoOOD sits after the backbone/refinement pipeline and after semantic occupancy prediction. The operational sequence is: image encoder $F^{3D}_{\text{refined}}$7 view transformation $F^{3D}_{\text{refined}}$8 coarse 3D features $F^{3D}_{\text{refined}}$9 PGSI $\hat{\mathbf{Y}}$0 PGTM $\hat{\mathbf{Y}}$1 refined 3D features $\hat{\mathbf{Y}}$2 $\hat{\mathbf{Y}}$3 occupancy head for semantic prediction, with EchoOOD consuming the refined features and predictions to produce a voxel-level anomaly map (Zhang et al., 1 Apr 2026).

The overall ProOOD overhead is small. The main text states that ProOOD introduces only 0.28M additional parameters overall. The supplementary efficiency analysis reports, for SGN, an increase from 28.16M to 28.44M parameters and from 522.34G to 580.66G FLOPs; for VoxDet, from 55.34M to 55.35M parameters and from 1311.10G to 1311.48G FLOPs. It further states that the core prototype learning components contribute only 0.01M parameters and most remaining overhead is due to an auxiliary occupancy head. Since EchoOOD itself is explicitly parameter-free at inference, its own cost is essentially the cosine-similarity, normalization, and local-prototype computations (Zhang et al., 1 Apr 2026).

EchoOOD is evaluated on the OOD benchmarks VAA-KITTI, VAA-KITTI-360, and VAA-STU, following OccOoD. The OOD metrics are AuPRC$\hat{\mathbf{Y}}$4 and AuROC, where AuPRC$\hat{\mathbf{Y}}$5 incorporates spatial tolerance via ground-truth dilation at 0.8 m, 1.0 m, and 1.2 m, while AuROC measures global separation quality (Zhang et al., 1 Apr 2026).

The clearest direct evidence appears in the anomaly-score ablation on VAA-KITTI. Compared with post-hoc baselines, EchoOOD achieves:

Method	AuPRC$\hat{\mathbf{Y}}$6 at 0.8 / 1.0 / 1.2 m	AuROC
Postpro	5.81 / 10.02 / 16.22	63.85
Energy	6.80 / 12.11 / 20.08	66.69
Entropy	7.38 / 12.91 / 21.05	66.24
ASS	14.47 / 20.96 / 29.65	64.20
EchoOOD	27.86 / 43.55 / 62.65	64.31

The paper summarizes this as improving AuPRC$\hat{\mathbf{Y}}$7 by more than 12% across all scales relative to other scoring methods while maintaining comparable AuROC. On the broader benchmark table, ProOOD with EchoOOD also outperforms ASS-based baselines by large margins across VAA-KITTI, VAA-KITTI-360, and VAA-STU. The paper additionally reports calibration-related gains for ProOOD(+SGN) on VAA-KITTI, reducing $\hat{\mathbf{Y}}$8 from 51.10 to 47.28, $\hat{\mathbf{Y}}$9 from 57.49 to 53.91, and $\mathbf{p}^{\mathrm g}_k$0 from 4.93 to 4.41, while increasing AuROC from 60.51 to 64.31 (Zhang et al., 1 Apr 2026).

Qualitatively, the paper highlights improved anomaly localization. One example states that SGN predicts the OOD object “clothing” as only a few scattered occupied voxels and assigns it a low anomaly score, whereas ProOOD reconstructs the occupancy better and assigns much higher anomaly scores. The supplement also remarks that VoxDet has “almost no OOD detection capability” despite strong in-distribution accuracy, and that adding ProOOD and EchoOOD “effectively mitigates this overconfidence tendency” (Zhang et al., 1 Apr 2026).

6. Scope, limitations, and terminological clarifications

EchoOOD is an OOD-scoring module for 3D semantic occupancy prediction, not a general-purpose “Echo”-branded method. Several unrelated papers in the provided literature use similar names but do not define EchoOOD. “AECMOS: A speech quality assessment metric for echo impairment” is a non-intrusive learned speech-quality metric for acoustic echo cancellation and is explicitly described as not being an out-of-distribution detector or uncertainty-aware failure detector (Purin et al., 2021). “Echoes: A semantically-aligned music deepfake detection dataset” is described as OOD-oriented in spirit for cross-dataset transfer, but the paper does not use the name EchoOOD (Pascu et al., 24 Mar 2026). “EchoXFlow” explicitly states that it does not mention EchoOOD anywhere and is best understood as a multimodal echocardiography dataset rather than an OOD benchmark (Stenhede et al., 6 May 2026).

Within its own domain, EchoOOD also has explicit limitations. The paper does not provide a clean component-wise ablation isolating only local logit coherence, only local prototype matching, only global prototype matching, or different fusion weights. It does not report a sensitivity analysis for the confidence threshold $\mathbf{p}^{\mathrm g}_k$1, similarity metric alternatives, or normalization choices. It does not specify the exact handling of missing or immature global prototypes at test time beyond the maturity-gating statement. These omissions constrain exact reproduction and make the relative contribution of each branch inferential rather than fully quantified (Zhang et al., 1 Apr 2026).

The method also depends on the underlying occupancy predictor. The supplement explicitly shows failure cases where small OOD objects are not reliably activated in the occupancy volume, leading to missed detections because EchoOOD scores occupancy features rather than raw sensor inputs directly. The paper further notes that extreme long-range or heavily occluded cases can produce missed anomalies or false positives in ambiguous background regions, and that the whole framework inherits the depth-estimation dependency of the underlying occupancy pipeline (Zhang et al., 1 Apr 2026).

Taken together, EchoOOD is best characterized as the “Unsupervised Local and Global Prototype Scoring” mechanism inside ProOOD: a training-free voxel-level anomaly scorer that detects semantic inconsistency across logits, local scene prototypes, and global class anchors. Its central methodological contribution is the fusion of scene-local and dataset-level prototype structure with class-wise logit coherence, tailored specifically to voxel-wise 3D occupancy prediction under long-tailed distributions (Zhang et al., 1 Apr 2026).