Question 1: Domain Generalization Papers

We chose RobustNet (2021)[6] and RuleAug (2023)[29] for our benchmark comparison because they represent two main DG strategies: Constraining the learned feature distributions and domain randomization. Although adding more recent DG techniques could enhance our analysis, many of these works do not provide pretrained models on Cityscapes or are based on different segmentation backbones, making direct evaluation and comparison challenging [25, d5].

To address the reviewer’s concern, we additionally evaluated a recently published work, CMFormer [d4] (2024), which adopts the Mask2Former architecture for DG in semantic segmentation. Using the provided pretrained model and official code, we performed inference on the ACDC-POC dataset and compared it with our method under the Mask2Former architecture. We note that the comparison is not entirely fair, as CMFormer uses a more powerful Swin Transformer backbone while we use ResNet50 following Mask2Anomaly. As shown in Table A2, CMFormer non-surprisingly performs better on known class segmentation with a 10-point gap in mIoU and 2 points in mAcc. However, its OOD segmentation performance is significantly worse, with over 60 points lower in AP and 30 points higher in FPR. This supports our finding that existing DG techniques might overly generalize to all types of distribution shifts, making it difficult to recognize unknown objects, raising safety concerns in autonomous driving scenarios.

Question 2: Validation

The main difference between domain generalization and domain adaptation techniques is that the former focuses on training one model to generalize to any data with different domain shifts, while the latter typically trains each model for each target domain. Our method belongs to domain generalization as we train one model on Cityscapes and evaluate it on various datasets with different domain shifts. We use the SMIYC validation set for model selection to ensure a fair comparison with other anomaly segmentation techniques in the benchmark, which also use the validation set for model selection.

[d4] Bi, Qi et al. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. AAAI, 2024. [d5] Li, Yumeng et al. Intra-Source Style Augmentation for Improved Domain Generalization. WACV, 2023.

We thank the reviewer for the time and detailed feedback. We address the SMIYC benchmark comparison in the general response and will revise the paper to correct the noted typos. Below, we address your specific concerns.

Weakness 2: Related Work & Novelty

Anomaly Segmentation: Our method falls under discriminative-based anomaly segmentation methods that utilize additional out-of-distribution (OOD) data to train models to distinguish between known and unknown data [3, 30, 21, 26]. We contribute to the construction of generative-based OOD data and training pipelines. The details of our OOD data generation are explained in the general response. Regarding training, previous works [3, 30, 26] directly train models using OOD score functions calculated on logits (e.g., MSP, Energy, Entropy). In contrast, our approach introduces a learnable uncertainty function, initially set as the standard OOD score function and optimized before fine-tuning the feature extractor for joint known class segmentation and OOD detection. This approach decouples the training process to mitigate competition and leverages the meaningful features learned from previous closed-world pretraining.

Domain Generalization: Existing methods for domain generalization include instance normalization or whitening [24, 6, 25] and domain-invariant feature learning through domain randomization [33, 29]. Our contribution lies in the latter category, where we propose a generative-based data augmentation technique. Unlike recent works that use generative models for feature [d1] or image randomization [d2], our technique concurrently generates data with both domain shifts and unknown objects. We believe that incorporating various distribution shifts is crucial to avoiding model biases and better handling domain and semantic shifts.

Weakness 3: Evaluation Metrics

We would like to clarify that we used two metrics, AP and FPR@95, in the presented SMIYC results. These metrics are widely recognized as the primary evaluation criteria in the anomaly segmentation literature [1]. It is a common practice in this field to present only the primary metrics in the main paper due to space constraints, as exemplified in [21, d3].

We agree with the reviewer that using multiple metrics provides a more comprehensive evaluation of the method. To address the reviewer’s concern, we have included the results of our method across all five metrics below. As shown in Table A1, our method consistently performs better or on par with RPL/Mask2Anomaly across all evaluation metrics. These additional results will be included in the appendix of the revised paper.

Table A1: Comparison of our methods with two recent works RPL and Mask2Anomaly on the SMIYC benchmark.

Weakness 4: Hyperparameters

We thank the reviewer for the suggestion. We have added results ablating the two types of hyperparameters in our method: selection ratio and loss margins. THe results shown in Table 1 of the attached PDF demonstrate the robustness of our model to various hyperparameters.

[d1] Gong, Rui, et al. Prompting diffusion representations for cross-domain semantic segmentation. BMVC (2024).
[d2] Jia, Yuru, et al. Domain-Generalizable Semantic Segmentation with Image Diffusion Models and Stylized Semantic Control. ECCV (2024).
[d3] Grcic et al., On Advantages of Mask-level Recognition for Outlier-aware Segmentation, CVPRW 2023.

Thank you for the thorough comments and many constructive suggestions. We appreciate the mention of the interesting work by Loiseau et al. [a1] and have discussed our differences, including [8] and [34], in the general response. We address the other concerns below.

1. Decouple Contribution

We thank the reviewer for the suggestions and have conducted additional experiments on Mask2Anomaly and RPL. The results are shown in Table C1. Below, we discuss the findings:

-Mask2Anomaly: We observe consistent improvements across all datasets, demonstrating the benefits of using our generated data. However, the final performance achieved by Mask2Anomaly + CG-Aug is still lower than the model fine-tuned with our pipeline. This highlights the efficacy of the proposed training design.

-RPL: The improvement is not as significant. This may be due to certain aspects of RPL's loss and training design being less suitable for our scenario. Firstly, RPL relies on the original network's predictions to supervise a learnable residual part. Since the original network does not generalize well to data with domain-shift, this results in imprecise supervision. Secondly, the RPL uncertainty loss focuses solely on increasing uncertainty for unknowns, without adequately addressing the known classes, particularly for augmented images. Additionally, restricting the trainable parameters to a residual block may limit the model's ability to learn more complex patterns, thereby reducing overall effectiveness.

Those results demonstrate that effectively utilizing the generated training data with multiple distribution shifts remains an open question. Our work takes a step towards analyzing the shortcomings of existing training designs, offering novel and effective strategies for better handling this data.

Table C1: We apply our coherent generative-based augmentation (CG-Aug) to the most recent works, RPL and Mask2Anomaly.

2. Contrastive Loss

We appreciate the reviewer’s kind suggestion and will add more related work in our model training section (3.3). Below, we discuss the novelty of the proposed contrastive loss.

Compared with contrastive loss in RPL: RPL employs their contrastive loss on a projected feature space, supervising OoD training with a combination of feature contrastive loss and an additional energy loss to maximize uncertainty scores for unknown class data. In contrast, we calculate our contrastive loss directly on the uncertainty scores. This direct supervision allows for more explicit and effective ranking of uncertainties across different data types. Experiments in Table C1 replacing our OOD loss with RPL's showed that RPL's feature contrastive loss alone barely improves performance. Even with combined losses, RPL's performance falls short of ours, demonstrating the efficacy of our direct supervision on uncertainty scores.

Compared with other OOD Losses: Existing OOD losses either maximize uncertainty scores solely for unknown data or supervise unknown and known data separately. In contrast, our method supervises the relative distance between OOD and inlier samples. We find this contrastive term to be more robust to hyperparameters and easier to optimize. To demonstrate its efficacy, we replaced the distance-based supervision with value-based supervision, similar to that used in Mask2Anomaly and PEBAL. As shown in Table 3, row #3, this change led to a performance decrease, validating the effectiveness of our loss design.

Table C2: Ablation study of our contrastive loss (In DeepLab v3+). Our loss performs better in both OOD detection and known class segmentation results.

4.Generation Failures

We agree with the reviewer that generative models still have important limitations, and appreciate the reviewer’s detailed examination and careful thought for generation failure cases and impact. Below, we first discuss how we currently deal with generation failures, followed by our analysis of the generation failure pattern, and examine how these failures affect model training.

• Noise-aware Learning: During training, we mitigate the impact of generation failures by using a sample selection strategy (Sec. 3.3.2). We note that for failure cases mentioned by the reviewer, some can be prevented via pixel selection during training. To better illustrate this, we include the loss and selection map corresponding to the images from Fig 5. in the attached PDF.

• Generation Failure Cases: Inspired by the reviewer, we examine our generated image and observe that generation failures typically occur in the following scenarios: (a) Remote scenes. (b) Small objects. (c) Text-related elements. This demonstrates the limitations of the current generative model and we hope they can be addressed in future research.

• Impact of Generation Failures:

  • Impact on Category Learning: Generation failures may adversely affect specific classes. We evaluated per-class segmentation results and compared them with the baseline model. Results are presented in Table C3. We find performance in six categories, such as fence, pole, and traffic sign, remains similar (differences less than 1%); and performance on vegetation is worsened by 3%, likely due to poor generation quality for this class.
  • Performance Saturation: We observe that performance tends to saturate with increasing generative data. Experiments varying the dataset scale from 1.0x to 2.0x and 3.0x Cityscapes sizes, as shown in Figure 1(b) in the supplementary material, suggest this saturation. This may be due to an interplay between the benefits of additional data and the negative impact of generation failures

We will include more examples of failure cases and a comprehensive discussion of the limitations of generative models in the revised paper and appendix.

Table C4. Per-class segmentation results in mIoU on ACDC-POC dataset.

Thank you for your constructive and detailed feedback.

• Regarding the comparisons: We appreciate your suggestion and will include RbA and M2F-EAM (Grcic et al.) in the revised manuscript. Thank you also for pointing out our typo in the general response, where (the second) Table A1 should be Table A2.

• Regarding the novelty: Thank you for acknowledging the contributions of our work. We are pleased that our additional experiments have clarified the advantage of our training pipeline, which is tailored to our specific task and effectively leverages our generated data.

  • Further experiments with COCO data as OOD: In response to your further suggestion, we have conducted experiments using COCO data as OOD to assess the efficacy of our proposed training pipeline without the generated data. We note that some training components, such as noise-aware learning and relative contrastive loss, are tailored for scenarios involving the generated domain-shift data. Therefore, in this evaluation, we focus primarily on assessing the effectiveness of our two-stage learnable uncertainty function and the relative contrastive loss applied between the original data and OOD data. The results in Table C5 show that our training method surpasses RPL and Mask2Anomaly on most metrics, and the combination with our generative data yields even greater performance improvements.
  • Comparison with POC: We appreciate your recognition of the novelty of our generative pipeline. Our work primarily addresses scenarios where both domain shifts and semantic shifts are present, which we believe are common in real-world applications. To address your concerns regarding performance under smaller domain shifts, we have included results on the FS LostAndFound and FS Static datasets, comparing our method with POC using the Mask2Anomaly training pipeline. As shown in Table C6, our method demonstrates superior performance on FS LostAndFound, indicating that our generated data closely resembles real OOD scenarios. However, on the FS Static dataset, our performance is lower than that of POC. We note that the FS Static dataset's OOD data is generated using a cut-and-paste technique, which may not reflect real-world OOD distributions. This could explain why simple COCO-pasted data achieves the best results on this dataset.

  Overall, we thank you again for your valuable input, which has helped us clarify the components of our design. We will incorporate all these discussions into the revised paper.

  Table C5: Performance of our training pipeline using COCO data, compared to previous methods Mask2Anomaly and RPL.

  Backbone	FT	OOD Data	Road Anomaly			SMIYC -RA21 (val)		SMIYC -RO21 (Val)
  			AUC$\uparrow$	AP$\uparrow$	FPR$\downarrow$	AP$\uparrow$	FPR$\downarrow$	AP$\uparrow$	FPR$\downarrow$
  Mask2Former	Mask2Anomaly	COCO	-	79.70	13.45	94.50	3.30	88.60	0.30
  	Ours	COCO	95.83	80.94	29.12	97.41	1.60	92.89	0.50
  	Ours	Ours	97.94	90.17	7.54	97.31	1.04	93.24	0.14
  DeepLabv3+	RPL	COCO	95.72	71.61	17.74	88.55	7.18	96.91	0.09
  	Ours	COCO	96.26	76.01	17.44	91.40	6.80	96.77	0.12
  	Ours	Ours	96.40	74.60	16.08	93.82	3.94	95.20	0.19

  Table C6: Comparison of our CG-Augmentation with POC on FS LostAndFound and FS Static, integrated into the Mask2Anomaly training pipeline. COCO results are from the official Mask2Anomaly paper; POC alt. and POC c. results are from the official POC publication.

  	FS_LostAndFound (val)		FS_Static (val)
  Mask2Anomlay+	AP$\uparrow$	FPR$\downarrow$	AP$\uparrow$	FPR$\downarrow$
  COCO	69.41	9.46	90.54	1.98
  POC alt.	68.8	11.4	87.4	3.1
  POC c.	73.0	9.2	87.0	2.1
  CGAug (ours)	76.56	10.17	85.70	7.16

Thank you for the time and constructive feedback. We discuss the novelty of our generative-based data augmentation and additional comparison results on SMIYC in the general response. We address your other concerns below.

Weakness 1: First Contribution

We thank the reviewer for highlighting the works by Zendel et al. [a] and Bevandic et al. [b]. These studies indeed demonstrate the necessity and feasibility of handling both semantic segmentation under domain shifts and anomaly segmentation by proposing datasets and baselines. However, our work differs in problem setting and method design. Below, we detail these differences and will revise the contribution description and related work section to better highlight our contributions.

• Our work builds upon these foundations by delving deeper into the core challenges of simultaneously enhancing model performance in both areas. Specifically, we provide a novel analysis of the limitations of Domain Generalization techniques in identifying anomaly objects, demonstrating the problem of 'over-generalization' for unknown regions. We also address the challenges faced by current state-of-the-art anomaly segmentation techniques, which often make errors in distinguishing between known objects with covariate shifts and real novel/anomaly objects (cf. Fig. 1). Furthermore, our findings indicate that simply combining techniques from both domains does not always yield optimal results for jointly handling domain shifts and semantic shifts. This is due to their focus on different levels (e.g., image-level domain shifts or object-level semantic shifts), leaving the challenge of distinguishing object-level domain shifts and semantic shifts unresolved (cf. Sec. 4.4).
• We also note that while [a] and [b] are seminal works in domain shifts and anomaly segmentation, their problem settings are still in early stages. For example, WildDash [a] contains very few image-level anomaly samples, and while [b] introduces object-level anomaly samples via cut-and-paste, it is limited to animals and involves significant artifacts. More recent benchmarks, such as SegmentMeIfYouCan and MUAD, offer domain and semantic shifts that better reflect real-world scenarios. We have tested existing OOD detection and common domain generalization methods on these benchmarks, highlighting their limitations.

Overall, we believe that our work takes a step further in addressing the challenges of jointly handling the two distribution shifts and filling a literature gap. We expect that our results will advocate for more research into developing algorithms that can improve both generalization and anomaly detection. We appreciate the reviewer's suggestion and will incorporate a discussion of these works in the related work section to better illustrate the contributions of our study.

Questions:

For the calculation of our relative contrastive loss (Eq. 4), we randomly sample an equal number of pixels from the unknown-class set, original known-class set, and augmented known-class set, to calculate the contrastive losses between unknown and (original or augmented) known pixels. The third contrastive loss term is directly calculated for all paired pixels from original and augmented images.

Limitations:

We discussed this limitation of our method in the conclusion. Similarly to what the reviewer mentioned here, the proposed augmentation strategy could be impacted by the quality of the generative model.

I find the rebuttal to be thorough enough to increase my score to borderline accept.

Thank you for your time and constructive feedback. We discuss the novelty of our proposed generative-based augmentation in the general response. We address your other concerns below.

Weakness 2: Hyperparameters for Relative Contrastive Loss

Our relative contrastive loss includes three terms, each with a margin value controlling the distance penalty limits. These margins are set based on the average uncertainty scores from the training set. Specifically, we compute the differences in uncertainty scores between unknown vs. original known data, unknown vs. augmented known data, original known vs. augmented known data, and set the differences as margins for these distance respectively. A histogram of uncertainty scores is provided in the Figure 1 (c) attached PDF for reference.

Moreover, our two-stage training framework first trains the uncertainty function based on the existing model, allowing this function to adapt to different scales. This provides flexibility in parameter setting even without prior knowledge. The experimental results, shown in Table 1 of the attached PDF, demonstrate the model's robustness across a wide range of hyperparameter variations. Specifically, maintaining the loss values within the same order of magnitude ensures that variations in parameters do not significantly affect the results.

Weakness 3: Two-Stage

Thanks for pointing this out. The motivation and benefits of our two-stage training design are clarified below.

• We observe that an initialized model with closed-world pre-training achieves a good feature representation while the initial uncertainty function, such as the energy score, is typically sub-optimal. Directly training the feature extractor with a sub-optimal uncertainty function risks disrupting the well-learned feature representations, which can harm known class segmentation and OOD detection (cf. Figure 4 (a)).

• Our two-stage training approach addresses this challenge by first optimizing the uncertainty mapping head based on the current feature representations. This allows the subsequent fine-tuning of the feature extractor to be more effective and less disruptive. Additionally, having a well-initialized uncertainty head before joint training helps minimize task competition between known class segmentation and OOD detection.

• In Figure 4(a), we empirically compare our two-stage training with single-stage training, showing that our approach (noted as 'second stage') outperforms the 'single stage' by nearly 10 points in AP. With a closer look, our first-stage training, which involves training only the uncertainty function already significantly improves the baseline model's performance (from 45% to 85% in AP). This demonstrates the large performance gap between different uncertainty functions under the same feature extractor.

Weakness 4: Conditioned Label Mask

The term "conditioned label masks" in the context of Line 208 refers to the label masks used to generate images. As explained in Eq. (1), these masks are created by cut-and-pasting the masks of novel objects onto the original training labels.

Weakness 5: Noise-aware training

• Sample Selection Mechanism: We use the 'small loss' criterion to determine whether a pixel has a clean label, a simple and widely used technique in the noisy label learning literature [b1]. Specifically, we calculate and rank the cross-entropy loss for each pixel. Pixels with smaller losses are selected for backpropagation, while those with larger losses are ignored (cf. Eq. (5-6)). This is effective because during training, a model first learns simple and clean patterns before fitting noisy data [b2]; We will include explanations and relevant literature in the paper to aid understanding. A visualization of our sample selection results is illustrated in Figure 4 (b) of the paper.

• Determining the Selection Ratio α: We determine α by visualizing the selection map of a small batch of data under several choices to ensure that visibly incorrect patterns are removed. To address the reviewer's concern, we conducted additional experiments with selection ratios ranging from 0.6 to 0.9, as detailed in Figure 1 (a) of the supplementary PDF. The results show that while including too many pixels (1.0) introduces noise, and including too few (0.6) removes useful regions, the model performance is stable within a wide range (0.7 to 0.9), demonstrating the robustness of the model to this hyperparameter.

Weakness 6: Hyperparameters & Applicability

We have demonstrated robustness to loss margins and selection ratio in the previous response, supporting the practicality of our approach for real-world applications.

[b1] Jiang, Lu, et al. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. ICML, 2018.
[b2] Arpit, Devansh, et al. A closer look at memorization in deep networks. ICML, 2017.

Thank you for your reply and for acknowledging our problem setting and proposed method.

Regarding your concern about hyperparameter tuning, as detailed in our responses to weaknesses 2 and 5, the hyperparameters in our method, including loss margins and selection ratios, can be set directly based on training data statistics without the need for further tuning. Additionally, in the attached PDF, we demonstrate the robustness of our method across a wide range of these hyperparameters. Other normal training hyperparameters, like the learning rate, are tuned once per backbone and then kept consistent throughout experiments. We want to emphasize that our method does not require more hyperparameter tuning than previous approaches like RPL and Mask2Anomaly. Developing an OOD training strategy that eliminates the need for hyperparameter tuning is an interesting direction, and we will leave it to future work.