ALOOD: Aligned LiDAR OOD Detection

• ALOOD is a language-driven method that aligns LiDAR object features with CLIP text embeddings, redefining OOD detection as a zero-shot similarity problem against known classes.
• It employs a lightweight CNN adapter and fuses local and global scene context with box geometry encoding to create robust LiDAR representations.
• Empirical results on nuScenes demonstrate competitive performance, with norm-scaled cosine similarity enhancing discrimination between in-distribution and OOD objects.

ALOOD denotes Aligned LiDAR representations for Out-Of-Distribution Detection, a post-hoc method for LiDAR-based 3D object detection that addresses the tendency of closed-set detectors to assign overly confident predictions to objects from unknown categories. Its defining mechanism is the alignment of LiDAR object features from a frozen 3D detector to the text-embedding space of a vision-LLM, so that OOD detection becomes a zero-shot similarity problem over known class names rather than a confidence heuristic over detector logits (Kösel et al., 9 Mar 2026).

1. Problem setting and conceptual scope

ALOOD is formulated for LiDAR-based 3D object detection in autonomous driving. A point cloud is represented as $p \in \mathbb{R}^{N \times 4}$, and a detector predicts $M$ 3D boxes $B=\{\mathbf{b}_j\}_{j=1}^M$, with each box $$\mathbf{b}_j=(x_c,y_c,z_c,l,w,h,\theta)\in\mathbb{R}^7$$. The detector is trained on a set of $K$ in-distribution classes, while OOD objects are categories absent from training but present at deployment (Kösel et al., 9 Mar 2026).

The method treats OOD detection as a post-hoc problem. The base detector is kept frozen, no OOD training samples are used, and no CLIP image features are required. Instead, ALOOD uses a frozen CLIP text encoder and aligns LiDAR object features into that language space. Once this alignment is learned, each detected object can be compared with cached text embeddings of the known classes; high similarity supports an ID decision, while low similarity supports an OOD decision (Kösel et al., 9 Mar 2026).

This formulation is evaluated on the nuScenes OOD benchmark, where training uses only scenes without OOD objects and validation/test contain 9 void classes treated as OOD. In that setting, ALOOD reframes OOD detection as zero-shot semantic matching to known-category text representations (Kösel et al., 9 Mar 2026).

2. Representation pipeline and language alignment

ALOOD is built on top of a frozen CenterPoint detector, with experiments on both voxel-based and pillar-based variants. Let the detector neck output a BEV feature map $\mathbf{F}\in\mathbb{R}^{C\times H\times W}$. A lightweight CNN adapter is applied to this map; it consists of two $3\times 3$ convolutions, ReLU activations, batch normalization, and a residual connection, yielding an adapted feature map $\mathbf{F}'\in\mathbb{R}^{C\times H\times W}$ (Kösel et al., 9 Mar 2026).

For each object, ALOOD extracts a local feature $\mathbf{f}_j\in\mathbb{R}^C$ by center pooling from $\mathbf{F}'$. It also computes a global scene descriptor

$M$0

and mixes local and global context as

$M$1

with $M$2 in the experiments. Box geometry is encoded separately through

$M$3

and the final LiDAR object representation is

$M$4

This representation is then projected into CLIP text space with a single linear layer,

$M$5

The paper reports that this linear alignment outperforms MLP alternatives (Kösel et al., 9 Mar 2026).

Text supervision is generated from prompts processed by a frozen CLIP text encoder. Two prompt forms are used during training:

• Simple prompt: “This object is a [cls].”
• Spatial prompt: “This object is a [cls] located at ([x,y,z]), with dimensions ([w,l,h]) and orientation [yaw] rad.”

Each training object is randomly assigned one of these prompts. The simple prompt supplies class semantics; the spatial prompt injects instance-specific geometry and is intended to improve generalization (Kösel et al., 9 Mar 2026).

3. Objective, scoring rule, and OOD decision

Similarity between aligned LiDAR features and text embeddings is computed by cosine similarity,

$M$6

Because multiple objects in a batch can share the same class, ALOOD uses a multi-positive contrastive objective rather than a one-positive CLIP loss. With $M$7 objects in a batch, class labels $M$8, positive set $M$9, and learnable temperature $B=\{\mathbf{b}_j\}_{j=1}^M$0, the alignment loss is

$B=\{\mathbf{b}_j\}_{j=1}^M$1

Training uses only ID data, ground-truth boxes, and the frozen text encoder; the base CenterPoint detector remains fixed throughout (Kösel et al., 9 Mar 2026).

At inference, ALOOD precomputes one text embedding per known class using the prompt “This object is a [cls].” For a detected object, it evaluates similarities

$B=\{\mathbf{b}_j\}_{j=1}^M$2

The predicted ID class is the class with maximal similarity. For OOD detection, the paper’s main score is the maximum similarity scaled by the aligned feature norm: $B=\{\mathbf{b}_j\}_{j=1}^M$3 An object is declared ID if $B=\{\mathbf{b}_j\}_{j=1}^M$4 and OOD otherwise, with $B=\{\mathbf{b}_j\}_{j=1}^M$5 chosen to satisfy a target ID true-positive rate such as $B=\{\mathbf{b}_j\}_{j=1}^M$6 (Kösel et al., 9 Mar 2026).

The use of norm scaling is empirically important. The paper reports that raw maximum similarity is less discriminative than the norm-scaled version, and that scaled MaxLogit in similarity space is the best-performing scoring rule among the tested options (Kösel et al., 9 Mar 2026).

4. Empirical results on nuScenes OOD detection

Evaluation uses FPR-95, AUROC, AUPR-S, and AUPR-E on the nuScenes OOD benchmark. The comparisons include Default Score, MSP, ODIN, MaxLogit, Energy, and a LiDAR-specific Rescaling baseline based on synthetic OOD generation (Kösel et al., 9 Mar 2026).

The main results are as follows.

Detector setting	Strongest comparison in paper	ALOOD
Voxel CenterPoint	Rescaling: FPR-95 36.96, AUROC 88.96	FPR-95 37.26, AUROC 90.15
Pillar CenterPoint	MSP: FPR-95 45.99, AUROC 90.44	FPR-95 38.78, AUROC 91.18

For voxel-based CenterPoint, ALOOD attains AUROC 90.15, AUPR-S 99.81, and AUPR-E 21.52, while Rescaling has slightly lower FPR-95 at 36.96 versus 37.26 for ALOOD. For pillar-based CenterPoint, ALOOD is best on all four reported metrics: FPR-95 38.78, AUROC 91.18, AUPR-S 99.72, and AUPR-E 24.66 (Kösel et al., 9 Mar 2026).

These results indicate that ALOOD is competitive with prior LiDAR OOD detectors on voxel-based CenterPoint and clearly superior on pillar-based CenterPoint. Because the detector itself is frozen, the method does not alter the standard detection accuracy of the underlying CenterPoint model (Kösel et al., 9 Mar 2026).

5. Ablations and design sensitivities

The ablation studies isolate several components of the method. First, the alignment head matters: a single linear projection is best. On voxel-based CenterPoint, linear alignment yields FPR-95 37.26 and AUROC 90.15, while an MLP with ReLU degrades to 44.86 and 88.27, and an MLP without ReLU gives 42.50 and 89.32 (Kösel et al., 9 Mar 2026).

Second, feature composition is not symmetric. Adding the CNN alone hurts, but the combination of CNN, global context, and box features is best. The strongest gain comes from box geometry. The full configuration reaches FPR-95 37.26 and AUPR-E 21.52, whereas using no CNN, no global context, and no box features gives 46.20 and 12.89 (Kösel et al., 9 Mar 2026).

Third, prompt design affects performance. Spatial prompts improve results over class-only prompts, and among prompt templates the paper reports that “This object is a [cls] …” is better overall than “A photo of a [cls] …”. On voxel-based CenterPoint, the latter yields FPR-95 38.01 and AUROC 89.62, whereas “This object is a [cls] …” reaches 37.26 and 90.15 (Kösel et al., 9 Mar 2026).

Fourth, the OOD score itself matters. The paper compares MSP, MaxLogit, and Energy in the aligned similarity space, with and without norm scaling, and concludes that including the aligned feature norm consistently improves separation. This supports the view that the norm of the aligned LiDAR representation carries additional ID/OOD information beyond cosine similarity alone (Kösel et al., 9 Mar 2026).

6. Position within the broader OOD literature

ALOOD belongs to a broader movement that replaces closed-set detector confidence with structurally richer post-hoc OOD signals, but its specific mechanism is distinctive: it aligns LiDAR features to a language space and performs zero-shot rejection against known class names (Kösel et al., 9 Mar 2026). That contrasts with several neighboring lines of work.

Adaptive test-time OOD detection is exemplified by AdaODD, which wraps a detector with a memory-based weighted KNN rule and adapts online by adding pseudo-ID and pseudo-OOD test features to memory; its core claim is that OOD detection should adapt to the test stream rather than use a fixed score forever (Zhang et al., 2023). ALOOD does not adapt online; it stays post-hoc and detector-frozen.

Gradient/prototype OOD detection is exemplified by GROOD, which constructs class prototypes and an artificial OOD prototype, then scores samples by nearest-neighbor distance in a gradient space derived from a prototype-based loss (ElAraby et al., 2023). ALOOD instead avoids synthetic OOD anchors and uses language embeddings as semantic targets.

Training-time ID-only energy variants such as AbeT modify the backbone itself by learning an input-dependent temperature and using an ablated learned-temperature energy score, with no OOD data and no test-time backward pass (LeVine et al., 2024). ALOOD differs in being explicitly post-hoc on a frozen 3D detector rather than a retrained classifier.

Few-shot prompt-based OOD detection is represented by AMCN, which also uses CLIP and textual prompts, but does so for few-shot image classification with learnable ID/OOD prompts and class-wise adaptive thresholds (Fang et al., 21 Jun 2025). ALOOD shares the use of language priors yet targets LiDAR 3D detection rather than few-shot image classification.

A separate branch of literature uses “ALOOD-style” reasoning in open-world active learning rather than test-time rejection. ALOE prioritizes diverse, likely unknown clusters for annotation in long-tailed image classification (Xie et al., 2024), while Joda separates known ID, discoverable novel classes, and irrelevant OOD in pool-based active learning through shared feature-space training, energy filtering, and SISOMe-based acquisition (Schmidt et al., 4 Mar 2025). This suggests that the acronym has acquired informal contextual use beyond the exact LiDAR method, but the explicit paper-defined term ALOOD refers to the language-aligned LiDAR OOD detector (Kösel et al., 9 Mar 2026).

7. Limitations and prospective extensions

ALOOD has several limitations that are directly visible from its formulation and experiments. It depends on a nontrivial modality alignment between LiDAR geometry and CLIP text embeddings; prompt wording and the inclusion of spatial fields materially affect results, which indicates sensitivity to prompt engineering (Kösel et al., 9 Mar 2026). It is also evaluated only on CenterPoint, and the paper itself identifies other detector families, especially two-stage 3D detectors, as future targets (Kösel et al., 9 Mar 2026).

Its performance is not uniformly dominant on every metric. On voxel-based CenterPoint, the Rescaling baseline slightly outperforms ALOOD on FPR-95 and AUPR-E, even though ALOOD is better on AUROC and AUPR-S (Kösel et al., 9 Mar 2026). The paper attributes part of this to benchmark characteristics, including the prevalence of many small OOD objects.

ALOOD also performs rejection, not open-vocabulary recognition. Unknown objects are flagged as OOD when similarity to all known text prototypes is low, but the method does not attempt to name the unknown category (Kösel et al., 9 Mar 2026). A plausible implication is that future extensions could combine ALOOD’s language alignment with broader score fusion strategies, since other domain-specific OOD frameworks such as COOD report large gains by combining complementary measures rather than relying on a single score family (Hogeweg et al., 2024).

Taken as a whole, ALOOD marks a specific shift in LiDAR OOD detection: from detector-confidence heuristics and synthetic outlier engineering toward language-conditioned semantic alignment. Its central claim is not merely that text embeddings are useful auxiliary features, but that they provide a workable target space in which LiDAR OOD detection can be posed as zero-shot similarity to known concepts (Kösel et al., 9 Mar 2026).