acceptodds
Under review as a conference paper at ICLR 2027

Out-of-Distribution Object Detection in Street Scenes via Synthetic Outlier Exposure and Transfer Learning

Abstract

Out-of-distribution (OOD) objects are often silently overlooked by object detectors, which produce no detection for them, leaving downstream modules with nothing to act on and no indication that anything was missed. A reliable object detector must detect OOD objects by localizing and correctly classifying them as OOD. Existing approaches primarily focus on OOD detection for image classification or rely on auxiliary branches, and typically do not treat in-distribution (ID) and OOD objects in a unified manner. We present SynOE-OD, a Synthetic Outlier-Exposure-based Object Detection method, which uses text-guided inpainting and an open-vocabulary object detector to generate and verify synthetic outliers offline. These are then used to fine-tune an object detector of choice with an added OOD class, which requires no auxiliary branch at inference time. On the OoDIS benchmark, SynOE-OD outperforms the strongest open-vocabulary baseline, reaching AP of  %,  % and  % on RoadObstacle, RoadAnomaly and Fishyscapes, against , and . This state-of-the-art performance holds for both fine-tuned closed-set and open-vocabulary detectors while maintaining competitive ID performance in street scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
out-of-distribution detection, outlier exposure, object detection, autonomous driving

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Fw1xOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper formulates OOD object detection as (n+1)-way detection. Synthetic outliers are generated by text-guided inpainting, verified and re-localized with an open-vocabulary detector, and used to fine-tune a detector with an additional OOD class. On OoDIS, the method reports AP50:95 of 57.6/29.4/39.2 on RoadObstacle/RoadAnomaly/Fishyscapes, compared with 48.5/22.7/22.7 for ZS-OVOD, and ablates vocabulary overlap, outlier proportion, annotation refinement, and ID-data diversity.

Critical strengths
  • The formulation is simple and operationally attractive. Treating OOD as an additional class removes the need for an auxiliary branch at inference and avoids dependence on real OOD data, while remaining compatible with both closed-set and open-vocabulary detectors.
  • The annotation refinement procedure is the most substantive technical contribution. Box refinement, label reassignment, and removal of empty inpaintings address a failure mode that is easy to overlook: without refinement, OOD AP falls to 1.3/4.2/1.9, compared with 46.9/24.4/29.3 after refinement. The 1,307 reassigned car instances provide a concrete estimate of how often inpainting preserves the original ID object.
  • The ablation suite is broader than is typical for this literature. The vocabulary-overlap experiment includes a matched non-overlapping control, and the outlier-proportion, per-class ID trade-off, and domain-diversity analyses each isolate a design choice.
Critical weaknesses
  • The comparison set is incomplete for a state-of-the-art claim. SIREN, SAFE, Unknown Sniffer, SyncOOD, and Dream-Box are discussed as related work but are not evaluated on OoDIS, despite being the closest trained OOD-detection methods. Without them, the empirical contribution cannot be positioned relative to the current literature.
  • The headline comparison conflates training regimes. ZS-OVOD is zero-shot, whereas SynOE-OD is fine-tuned on NuImages and optionally Cityscapes or BDD. The abstract's 57.6 versus 48.5 therefore does not isolate the effect of synthetic outlier exposure. The appropriate control—ID-only fine-tuning—is reported in Table 2 and the 0% row of Table 8, but it is not foregrounded.
  • The operating point and model-selection protocol are not adequately controlled. On RoadAnomaly, Table 2 reports 419 true positives against 1,222 false positives, implying precision of approximately 26%; yet the paper dismisses FPR95 and AUROC as inapplicable and does not report false positives per image or a threshold sweep. The OOD class name and training configuration were also selected using the OoDIS test sets, and several cells differ across tables (48.8, 48.7, and 47.6 for the same setting), which further weakens the reported comparison.
Questions for authors

NA

2 comments

We thank the reviewer for the thorough review and the precise comments. We answer them point by point.

W1: Incomplete comparison set, missing SIREN, SAFE, Unknown Sniffer, SyncOOD, and Dream-Box.

We agree that these comparisons were missing. We extended the comparison set and now evaluate Unknown Sniffer, SIREN, Dream-Box, SAFE and SyncOOD. We also included a copy-paste baseline proposed by reviewer 1.

MethodRoadObstacleRoadAnomalyLostAndFound
Vanilla Faster R-CNN R501.772.043.32
+ SAFE1.650.892.24
+ SIREN1.372.012.26
+ Unknown Sniffer3.711.814.84
+ Dream-Box0.731.692.34
GDINO Swin-T (zero-shot)48.319.526.4
+ SyncOOD36.09.417.9
DINO-DETR R50 (Paper Table 8, 0% outliers)12.02.45.2
+ Copy-paste41.7 ± 0.521.3 ± 0.620.5 ± 2.2
+ SynOE-OD48.7 ± 0.126.0 ± 0.131.7 ± 1.2
Table R6. All methods are trained on NuImages except for ZS-GDINO, and evaluated identically. Each is listed under the detector it is built on. Copy-paste uses the same detector and the same (n+1)-class formulation as SynOE-OD, differing only in where the outliers come from.

Results: Of the five methods, four underperform the detector they are built on. SAFE keeps the detector frozen and trains a classifier on sensitivity-aware features, using perturbed in-distribution samples and scores the boxes the detector already produced. SIREN also leaves the detection stage unchanged for unknown objects and shapes the ID class representations and scores each box by its fit to that representation. SyncOOD likewise trains a classifier on boxes of a frozen detector. Therefore these methods underperform their base detector. Dream-Box trains the region proposal network, but on outliers that are regenerated ID objects with their original annotations and it also underperforms its base detector.

Unknown Sniffer learns a generalized object confidence score from known objects, and with negative energy suppression, keeps unknown objects from being discarded as background. It is the only method that exceeds its base detector in recovering objects, which is what it is designed to do. In addition, the Unknown Sniffer paper reports a performance drop on images that contain both known and unknown objects. This is the most common case in street scenes.

W2: The headline comparison conflates training regimes.

Thank you for this observation. We agree. The comparison in the abstract does not isolate the effect of synthetic outlier exposure. The correct control is the identically trained ID-only model. It reaches 11.3, compared with 57.6 for SynOE-OD. This is also below the ZS-OVOD baseline (48.5). We will foreground this comparison in the main results and add the ID-only model as a row to Table 1.

W3.1: The operating point and model-selection protocol are not adequately controlled. On RoadAnomaly, Table 2 reports 419 true positives against 1,222 false positives, implying precision of approximately 26%; yet the paper dismisses FPR95 and AUROC as inapplicable and does not report false positives per image or a threshold sweep.

Thank you for this comment. We agree that false positives per image and a threshold sweep were missing, and we add them below.

We would like to clarify why FPR95 and AUROC are not applicable in our setting. FPR95 and AUROC measure how well two sets of scores are separated: scores on ID data and scores on OOD data. An object that is not detected appears in neither set. These metrics therefore cannot penalize a missed detection, so AP is the right metric.

The benchmark returns the number of predictions per frame, (TP + FP) / #images. We report it for the three OoDIS tracks. For LostAndFound we report false positives per image and a precision-recall curve from a threshold sweep.

ModelRoadObstacleRoadAnomalyFishyscapes
SynOE-OD (DD:R50)2.2915.962.41
SynOE-OD (DD:Swin-L)2.0410.272.22
SynOE-OD (GD:Swin-T)3.0316.773.57
Table R7. Predictions per frame, (TP + FP) / number of images, on the three OoDIS tracks.

False positives (FP) per image on LostAndFound. Over LostAndFound scenes, SynOE-OD produces 2.13 FP per image, the ID-only 3.27, and VOS 0.77. SynOE-OD therefore detects more than twice as many objects as the ID-only control (1,245 against 553 true positives) with fewer false positives per image, while VOS's low rate comes with only 334 true positives.

Precision-recall curve on LostAndFound. SynOE-OD reaches a recall of 0.646, while the ID-only model and VOS stop at 0.287 and 0.173.

RecallSynOE-ODID-onlyVOS
0.10.950.330.25
0.20.870.25–
0.30.76––
0.40.62––
0.50.52––
0.60.40––
Table R8. Precision at fixed recall on LostAndFound.

W3.2: Class name, configuration, and differing cells

Thank you for checking the tables so carefully. This comment has three parts, and we answer them separately:

  • Class name. We would like to clarify that we did not select the class name on the test sets. SynOE-OD (GDINO:SwinT) is text-conditioned, so the class name must be a word that the text encoder can represent. "OOD" is split into meaningless tokens (Table 4a). "anomaly" is a single token. The training loss confirms this before any evaluation: it starts at a lower value. For SynOE-OD (DD:R50 and DD:Swin-L) the class name is irrelevant, because these models have no text encoder.

  • Training configuration. We selected the training configurations on LostAndFound, based on OOD performance and its balance with ID performance.

  • Numbers across tables. We agree that the paper does not make it clear enough, but they do not come from the same setting: 48.8 trained on NuImages and Cityscapes (domain-gap study), 48.7 trained on NuImages (standard setting), 47.6 schedule matched with the ID-only baseline so it is comparable to it in that table. Each table has one fixed setting and answers one fixed question. The numbers are therefore not meant to be compared across the tables. We will state the setting in each table caption clearly and use proper indicators to avoid confusion.

We hope that we sufficiently addressed the concerns raised, and we are open to further discussion. We thank the reviewer for their time.

Reviewer_SczwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method for OOD object detection in street scenes. They generate a set of classes unrelated to driving with ChatGPT, then inpaint these into existing street scenes with off-the-shelf diffusion models. Afterwards, their bounding boxes are obtained with an open-vocabulary object detector, and an OOD-aware object detector is trained by adding an additional class to which all OOD objects belong. The method is evaluated on three benchmarks and by a series of ablations.

Critical strengths
  • The results of the proposed method are strong.
  • The paper has figures that illustrate the points made, and the text is well-written.
  • The paper contains many ablation studies that add to the understanding.
Critical weaknesses

The VOS results are very low. While VOS could be worse than the proposed method, the fact that it is (far) worse than ID-only is strange. What are the settings used, and does performance stay low for different values of the important hyperparameters? Appendix C of the VOS paper shows these can have a strong effect. To me, this is the main point that needs to be addressed clearly.

Otherwise, the main weakness of the paper would be the novelty of the findings, as some of the main insights of this work can be found in previous works: papers such as NPOS and VOS show the benefit of synthetic outliers for OOD detection performance, follow-up methods such as Dream-OOD and B-OOD show that pixel-space outliers can perform better than feature-space outliers, and using a k+1th class as an OOD detector is a proven strategy in many papers like [1]. As a result, the novelty lies in the specific application to OOD object detection in street scenes, which is somewhat limited but valid, and I do not consider this a major issue.

[1] Vernekar, Sachin, et al. "Out-of-distribution detection in classifiers via generation." arXiv preprint arXiv:1910.04241 (2019).

Questions for authors
  • What settings are used for the VOS baseline? How where they selected? What happens if these are adjusted? Why are the results so much worse than ID-only?
  • The main results of Table 1 can also be extended. For instance, the related work mentions Dream-Box as tackling the same problem, can this be compared to?
  • There are no details on the manual review of the updated annotations.
  • What are the OOD detection results under identical ID performance?
  • Why are the training settings (e.g., datasets) different between tables 1 and 2?
  • There is a lack of discussion on computational complexity. The proposed training pipeline makes use of several large-scale models and requires a significant amount of compute, which should be clearly stated.

Small:

  • Make sure to use the right \cite command, the brackets around citations are missing.
2 comments

We thank the reviewer for the constructive feedback and the detailed questions.

W1: Explanation on VOS required. Q1: What settings are used for the VOS baseline? How where they selected? What happens if these are adjusted? Why are the results so much worse than ID-only?

Thank you for this important question. We used the default settings of the released VOS code. For the dataset-dependent parameters we used their BDD-100K configuration as it is also a driving dataset. We used a 12,000 iterations schedule on NuImages, scaled to our training set size. The hyperparameters in their Appendix C, as far as we understood, act on the virtual outliers and the energy loss which apply at the classification head. Therefore we did not tune or change them.

VOS is below our ID-only baseline because the two methods use different detectors: VOS is built on Faster R-CNN, our ID-only baseline is DINO-DETR, and our SynOE-OD is built on DINO-DETR. Compared with the vanilla Faster R-CNN, VOS performs comparably (Table 4 row 1 and 3), both trained on NuImages. VOS's energy score does not overcome the limit of the detector in finding unknown objects. Furthermore, VOS synthesizes its outliers from the ROI features in a two-stage detector, whereas DINO-DETR has neither an RPN nor ROI features.

MethodID (mAP)RORALF
VOS base detector (Faster R-CNN)421.772.043.32
ID-only (Paper Table 2)42.211.32.45.4
VOS421.42.22.1
Table R4. VOS, Faster R-CNN and ID-only (DINO-DETR), trained on NuImages. RO = RoadObstacle, RA = RoadAnomaly, LF = LostAndFound.

Q2: The main results of Table 1 can also be extended.

Thank you for this suggestion. We evaluated Dream-box and will add it to the paper.

Dream-Box: trains the region proposal network of Faster R-CNN on outliers. The classification and box regression heads are trained on ID classes only. The learned classifier on the energy of the classification scores identifies OOD objects. The outliers are diffusion-generated regenerations of ID objects that keep the original annotations.

  • Setting. We ran the released code for the full pipeline with one change. Their generator inpaints the whole frame at 512×512. This suits PASCAL VOC, but NuImages frames are 1600×900 and many objects would become very small. We therefore crop around each box.
  • Outlier count. Their generator produced 9,073 outlier instances. We compare against SynOE-OD trained with 50% of our outliers (about 7,800 of 15,604), fewer outliers than Dream-Box.
  • Result. Dream-Box is built on a two-stage detector, so it is compared against its own base detector, which it underperforms. | Method | Outliers | ID mAP | RO | RA | LF | |:-------|:--------:|:------:|:-------:|:-------:|:-------:| | Vanilla Faster R-CNN | none | 42 | 1.77 | 2.04 | 3.32 | | + Dream-Box | 9,073 | 40.5 | 0.73 | 1.69 | 2.34 | | SynOE-OD (DD:R50) | ≈ 7,800 | 39.9 | 48.2 | 26.2 | 30.4 | Table R5. Comparison with Dream-Box, trained on NuImages.

We also evaluated other baselines and will add them to the paper.

Q3: There are no details on the manual review of the updated annotations.

Thank you for pointing this out. The manual review was a qualitative inspection of generated samples during development. We did not log the number of inspected samples. We will state this clearly in the paper.

Q4: What are the OOD detection results under identical ID performance?

Our outlier-proportion ablation (Table 8) answers this. With 25% outliers, ID performance is almost identical to the 0% outliers setting (40.0 vs 40.4 ID mAP) and OOD detection results move to 46.4 / 26.1 / 29.9 from 12 / 2.4 / 5.2 on RO, RA, LF.

Q5: Why are the training settings (e.g., datasets) different between tables 1 and 2?

Our domain-gap ablation (Table 5) shows that combining datasets improves domain transfer for OOD detection. NuImages plus Cityscapes gave better performance on LostAndFound. We therefore used this combination for Table 1. Table 2 answers a different question: it compares SynOE-OD with VOS and the ID-only baseline. For a controlled comparison, all three methods must see the same ID data. VOS and the ID-only baseline are trained on NuImages, and so is SynOE-OD in Table 2. ID-only and SynOE-OD share the identical training schedule. We will state this in the paper.

Q6: There is a lack of discussion on computational complexity.

Thank you for raising this point. We agree and will report the costs explicitly and add this information to the paper. All experiments were conducted on 3 NVIDIA Quadro RTX 8000 (48 GB VRAM) GPUs. The data pipeline is a one-time offline cost of about 26 GPU hours (outlier generation plus annotation and refinement). Fine-tuning itself takes 5 GPU hours for DINO-DETR R50, 9 for GroundingDINO Swin-T and 18 for DINO-DETR Swin-L. At inference there is no additional cost.

We hope that we have sufficiently addressed the concerns raised, and we are open to further discussion. We thank the reviewer for their time.

Reviewer_2AUuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies out-of-distribution (OOD) object detection in street scenes, emphasizing a failure mode in which unknown objects are suppressed as background and are therefore never localized. The proposed SynOE-OD pipeline first constructs a vocabulary of unusual street-scene objects, uses Stable Diffusion inpainting to insert these objects into ID scenes, and employs Grounding DINO to refine bounding boxes and correct failed or mislabeled inpaintings. The resulting synthetic OOD instances are assigned a single additional OOD class, and standard closed-set or open-vocabulary object detectors are fine-tuned to jointly localize ID and OOD objects without an auxiliary OOD branch at inference. Experiments on OoDIS and LostAndFound with DINO-DETR and Grounding DINO show strong improvements over the reported baselines, while mostly preserving ID mAP. The ablations further show that annotation quality is crucial, whereas the exact outlier vocabulary and the total amount of synthetic data are less critical.

Critical strengths
  1. The task formulation is practically meaningful. The paper targets OOD objects that are completely missed by a standard detector, rather than only assigning an OOD score to boxes that the detector already proposed. Training the OOD class as a regular detection target is simple, compatible with both closed-set and open-vocabulary detectors, and avoids an extra inference-time branch.
  2. The empirical study is useful within the chosen setup. SynOE-OD obtains large gains on the OoDIS benchmark across multiple detectors and backbones, with only a modest drop in ID performance. The ablations on outlier proportion, semantic overlap, domain diversity, and annotation refinement are informative; in particular, the very large degradation without accurate box refinement highlights an important practical lesson for synthetic OOD supervision.
Critical weaknesses
  1. The methodological novelty appears limited and largely engineering-oriented relative to closely related work. The core recipe is a composition of existing components: diffusion-based object inpainting, LLM/open-vocabulary concept discovery, foundation-model-based annotation/refinement, outlier exposure, and fine-tuning a detector with one additional OOD class. This is particularly close in spirit to 'Can OOD Object Detectors Learn from Foundation Models?' (SyncOOD, ECCV 2024), which already uses foundation models to synthesize, curate, filter, and annotate context-aware OOD objects for detector training, and to 'Dream-Box: Object-wise Outlier Generation for Out-of-Distribution Detection' (CVPRW 2025), which explicitly studies diffusion-based object-wise pixel-space outlier generation for OOD object detection. The main distinction of SynOE-OD is to supervise synthetic outliers directly as an (n+1)-st detection class so that localization and OOD classification are learned jointly, plus a careful annotation-cleaning pipeline. I find this distinction useful, but currently closer to a training/data-engineering choice than a substantial new learning principle. The paper's own ablations reinforce this concern: performance is extremely sensitive to annotation refinement, while reducing the amount of synthetic data or removing semantically overlapping outlier classes causes much smaller changes. This suggests that the dominant contribution may be high-quality OOD box supervision rather than a new OOD detection mechanism.
  2. The comparison set is too narrow for a 2027 submission and does not establish the claimed state of the art against the most relevant recent literature. The main OoDIS table compares primarily against ZS-OVOD and UGainS; the additional training-based comparison includes VOS and an ID-only control. However, the closest synthetic-data methods, especially SyncOOD and Dream-Box, are discussed in related work but are not quantitatively compared under a matched detector/data/evaluation protocol. This omission is important because these methods address nearly the same question of using generated object-level outliers to improve OOD object detection. More broadly, recent 2025-2026 work has continued to refine the formulation and evaluation of unknown/open-set object detection, including 'Rethinking Open-Set Object Detection: Issues, A New Formulation, and Taxonomy' (IJCV 2025) and road-scene unknown-object methods such as 'A Driving Oriented Objectness score for detecting known and unknown objects in road scenes' (CVIU 2026). These works are not necessarily directly comparable under OoDIS AP, but their absence makes the empirical positioning look dated and incomplete. At minimum, the paper should provide matched reimplementations for the closest synthetic-OOD methods where feasible, and a much stronger set of simple controls. In particular, a real-object copy-paste baseline using external OOD objects with accurate boxes, followed by exactly the same (n+1)-class fine-tuning, is essential to determine whether Stable Diffusion + LLM + GDINO is actually necessary or whether the gain mainly comes from exposing the detector to any sufficiently diverse, correctly localized negative objects.
Questions for authors
  1. Can the authors add matched quantitative comparisons with the closest synthetic OOD object-detection methods, particularly SyncOOD ('Can OOD Object Detectors Learn from Foundation Models?', ECCV 2024) and Dream-Box ('Object-wise Outlier Generation for Out-of-Distribution Detection', CVPRW 2025)? If their original protocols are incompatible with OoDIS, please adapt the methods to the same base detector and report a controlled comparison, or clearly separate which component cannot be transferred. Demonstrating a consistent advantage under matched training data, detector, and evaluation would materially improve my assessment of both novelty and significance.
  2. Can the authors isolate whether the gains come from the specific generative pipeline or simply from having accurately localized OOD supervision? A key baseline would be to copy-paste real external objects (e.g., from an external detection dataset, excluding the ID label space) into the same street scenes and train exactly the same (n+1)-class detector. It would also be useful to compare Stable-Diffusion inpainting with simpler augmentation under a matched number of OOD boxes.
4 comments

We thank the reviewer for the constructive feedback and the valuable suggestions. The two weaknesses and the two questions raise the same concerns. We address them together point by point below and return to the remaining points at the end.

W1: Closeness to SyncOOD and Dream-Box in spirit. Q1: Can the authors add matched quantitative comparisons with the closest synthetic OOD object-detection methods?

Thank you for this request. We ran both comparisons.

SyncOOD is a lightweight plug-in. It trains an MLP on outliers and on box features of a frozen detector. The MLP scores boxes that the detector has already produced.

  • What cannot be transferred. The outlier generation pipeline of SyncOOD is not released. We therefore train their classifier on our outliers.
  • Controlled setting. Both methods use the same base detector (GDINO Swin-T), the same training data (NuImages) and the same synthetic outliers.
  • Result. We apply SyncOOD on top of zero-shot GDINO, so it operates on strong object proposals. SyncOOD lowers the AP of its base detector. It replaces the detector confidence with the output of its classifier, and it cannot recover objects that the detector missed. Whereas SynOE-OD changes which objects are detected, not only how existing boxes are ranked. | Method | RO | RA | LF | |:-----|:-------:|:-------:|:-------:| | GroundingDINO zero-shot (Paper Table 3) | 48.3 | 19.5 | 26.4 | | + SyncOOD | 36.0 | 9.4 | 17.9 | | SynOE-OD (GD:Swin-T) | 54.3 | 29.3 | 29.2 | Table R1. Comparison with SyncOOD. Base detector: GroundingDINO Swin-T. Training data: NuImages with our synthetic outliers. RO = RoadObstacle, RA = RoadAnomaly, LF = LostAndFound. All values are AP.

Dream-Box: Dream-Box trains the region proposal network (RPN) of Faster R-CNN on outliers. The classification and box regression heads are trained on ID classes only. The learned classifier on the energy of the classification scores identifies OOD objects. The outliers are diffusion-generated regenerations of ID objects that keep the original annotations.

  • Setting. We ran the released code for the full pipeline with one change. Their generator inpaints the whole frame at 512×512. This suits PASCAL VOC, but NuImages frames are 1600×900 and many objects would become very small. We therefore crop around each box. We used their generator at two of their noise levels (σ=\sigma = 0.01 and 2.5) and report the stronger result.

  • Outlier count. Their generator produced 9,073 outlier instances. We compare against SynOE-OD trained with 50% of our outliers (about 7,800 of 15,604), fewer than Dream-Box.

  • Result. Dream-Box is built on a two-stage detector and the method is the region proposal network (it trains the RPN on outliers), whereas DINO-DETR has no RPN. Therefore it is compared against its own base detector, which it underperforms. | Method | Outliers | ID mAP | RO | RA | LF | |:-------|:--------:|:------:|:-------:|:-------:|:-------:| | Vanilla Faster R-CNN | none | 42 | 1.77 | 2.04 | 3.32 | | + Dream-Box | 9,073 | 40.5 | 0.73 | 1.69 | 2.34 | | SynOE-OD (DD:R50) | ≈ 7,800 | 39.9 | 48.2 | 26.2 | 30.4 |

Table R2. Comparison with Dream-Box, trained on NuImages. Dream-Box is built on Faster R-CNN. SynOE-OD uses DINO-DETR R50 (DD:R50) and 50% of our outliers. RO = RoadObstacle, RA = RoadAnomaly, LF = LostAndFound.

W2: a real-object copy-paste baseline using external OOD objects with accurate boxes.

Q2: A key baseline would be to copy-paste real external objects into the same street scenes and train exactly the same (n+1)-class detector.

We built a copy-paste baseline as requested by the reviewer, that copies real external objects into the same street scenes and matches our generated set: 10,702 images, 82,197 ID annotations and 15,604 outliers. The outliers are segment-wise crops from 74 non-ID categories of COCO val2017. We paste them into the same regions/locations that our pipeline uses for generation. Model (DD:R50) and training settings are identical for both rows in the table below.

MethodRORALF
Copy-paste41.7 ± 0.521.3 ± 0.620.5 ± 2.2
SynOE-OD48.7 ± 0.126.0 ± 0.131.7 ± 1.2

Table R3. Copy-paste baseline and SynOE-OD. Both use DINO-DETR R50 (DD:R50), identical training settings, and both are reported at the same epoch and averaged over seeds. RO = RoadObstacle, RA = RoadAnomaly, LF = LostAndFound.

Copy-paste is a respectable baseline, but our SynOE-OD is clearly stronger. Accurately localized OOD supervision is crucial, as our annotation-refinement ablation shows. This experiment further shows that the quality, realism and diversity of the outliers matter as well. In addition, our generation is not limited to the vocabulary of existing annotated detection datasets. We thank the reviewer for this suggestion and will add this ablation to the main paper.

[W2] We thank the reviewer for pointing out two relevant works and will add them both to the related work.

We hope that we have sufficiently addressed the concerns raised, and we are open to further discussion. We thank the reviewer for their time.

The additional comparisons with SyncOOD and Dream-Box, along with the real-object copy-paste baseline, address some of my experimental concerns and strengthen the empirical evaluation.

However, my main concern regarding methodological novelty remains. The proposed approach still primarily integrates existing techniques, without introducing a substantial new learning mechanism. Moreover, the SyncOOD comparison does not include its original generation pipeline, while the Dream-Box comparison uses a different detector. The evaluation also remains insufficiently positioned against recent work on OOD and unknown object detection.

Overall, I acknowledge the additional experimental evidence and the practical value of this work, but I do not find the rebuttal sufficient to resolve my main concerns. Therefore, I will maintain my original rating of 2 (Weak Rejection).

We thank the reviewer for the follow-up. We would like to clarify how we answered the initial questions by citing them alongside our previous response:

Q1: Can the authors add matched quantitative comparisons with the closest synthetic OOD object-detection methods, particularly SyncOOD[1] and Dream-Box[2]? If their original protocols are incompatible with OoDIS, please adapt the methods to the same base detector and report a controlled comparison, or clearly separate which component cannot be transferred. Demonstrating a consistent advantage under matched training data, detector, and evaluation would materially improve my assessment of both novelty and significance.

(please adapt the methods to the same base detector)

SyncOOD is run on the same base detector (GroundingDINO Swin-T), the same training data (NuImages), the same outliers and the same evaluation: 36.0 / 9.4 / 17.9 against 54.3 / 29.3 / 29.2 (Table R1).

Dream-Box is compared against its own base detector, a vanilla Faster R-CNN on the same training data. We have now also trained our SynOE-OD (n+1) formulation on Faster R-CNN. Dream-box reaches 0.73 / 1.69 / 2.34 against its base detector's 1.77 / 2.04 / 3.32 (Table R2); our formulation reaches 41.6 / 18.1 / 14.2 on RoadObstacle, RoadAnomaly and LostAndFound.

(clearly separate which component cannot be transferred)

SyncOOD: The component that cannot be transferred is its generation pipeline, as it is not publicly released. Therefore we used ours instead. This cannot change the outcome as SyncOOD scores boxes produced by a frozen detector, therefore the boxes it can score are fixed.

Dream-Box: The component that cannot be transferred is its base detector due to its OOD detection mechanism. It depends on the region proposal network (RPN and ROI head), which has no counterpart in DINO-DETR.

W1, W2: ...The paper's own ablations reinforce this concern: performance is extremely sensitive to annotation refinement, while reducing the amount of synthetic data or removing semantically overlapping outlier classes causes much smaller changes. This suggests that the dominant contribution may be high-quality OOD box supervision rather than a new OOD detection mechanism. Q2: Can the authors isolate whether the gains come from the specific generative pipeline or simply from having accurately localized OOD supervision?

Copy-paste exactly tests this: high quality OOD box supervision and real objects with exact boxes, on the same detector formulation (Table R3 above). It does not reach SynOE-OD's performance. The results show that it is not only high quality boxes that account for the result: the generated outliers themselves, as done in our method, contribute significantly to the OOD detection performance.

Novelty: SynOE-OD changes the OOD detection learning mechanism: it changes the detection stage rather than re-ranking the already detected boxes. It learns to localize and classify unknown objects, which are otherwise missed by a standard detector as background, while jointly detecting the ID classes. No prior OOD detection method does this. The closely related methods SyncOOD and Dream-Box, use the same components but apply them differently, so their data generation pipelines differ substantially from ours. Furthermore, our SynOE-OD targets street scenes, where both ID and OOD objects appear in the same scene and the detector must separate them.

Q1: ... Demonstrating a consistent advantage under matched training data, detector, and evaluation would materially improve my assessment of both novelty and significance.

We have demonstrated a consistent advantage under matched training data, detector, and evaluation, in three requested comparisons by the reviewer: SyncOOD, Dream-Box and copy-paste.

Positioning against related work. In this rebuttal we evaluated SIREN, SAFE, SyncOOD, Unknown Sniffer, Dream-Box and a copy-paste control, and will add Hosoya et al. (IJCV 2025) and Bunel et al. (CVIU 2026) to related work. None of these methods reports results on OoDIS, so each required reimplementation. If there is specific further work the reviewer has in mind, we include it to related work.

[1] 'Can OOD Object Detectors Learn from Foundation Models?', ECCV 2024

[2] Object-wise Outlier Generation for Out-of-Distribution Detection', CVPRW 2025

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for all reviewers and the area chair

We thank all reviewers for their time and their constructive feedback. The reviewers (R1, R2, R3) name the following strengths of our work:

1. Relevant problem. (R1) SynOE-OD addresses unknown objects that are missed as background.

2. Practical task formulation. (R1, R3) One detector jointly detects ID and OOD objects in street scenes.

3. Simple method. (R1, R3) SynOE-OD during inference needs no auxiliary branch.

4. Strong results (R1, R2) on OOD object detection benchmarks with only a modest drop in ID results.

5. Broad ablations. (R1, R2, R3) R3 identifies annotation refinement as a substantive technical contribution.

The main concerns are missing comparisons. We address them with new experiments and will add these results to the paper:

  • comparisons with SyncOOD, Dream-Box, Unknown Sniffer, SAFE, SIREN,
  • a copy-paste baseline with real objects,
  • a Faster R-CNN comparison for VOS,
  • false positives per image and a precision-recall curve,
  • compute costs.

The reviewers suggested several additional baselines and we evaluate all of them. OOD object detection requires localizing and correctly classifying unknown objects. OoDIS benchmarks and LostAndFound evaluate this for street scenes. We compare the methods along two properties, which gives us three groups in the following table:

GroupMethodsOutputs a box for unknown objects?Prevents suppression as background?
1NPOSNo, image classificationNo
2VOS, SIREN, SAFE, SyncOODOnly if the detector already produced itNo
3Unknown Sniffer, Dream-Box, SynOE-OD (ours)Yes, changes the detection stageYes

Group 2 methods score boxes the detector has already produced, so an object that is missed remains overlooked. Group 3 changes the detection stage and can recover an object that would otherwise be missed. Within Group 3 the methods differ in how unknown objects are prevented from being suppressed: Unknown Sniffer generalizes objectness from ID objects and is not supervised with unknown objects; Dream-Box trains its proposal network on outliers that are regenerated ID objects and differs substantially in outlier generation from SynOE-OD; SynOE-OD is supervised with localized novel synthetic objects. The results follow this grouping: no prior method improves substantially on the detector it is built on but SynOE-OD does.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.