FreeMask: Synthetic Images with Dense Annotations Make Stronger Segmentation Models

Semantic segmentation has witnessed tremendous progress due to the proposal of various advanced network architectures. However, they are extremely hungry for delicate annotations to train, and the acquisition is laborious and unaffordable. Therefore, we present FreeMask in this work, which resorts to synthetic images from generative models to ease the burden of both data collection and annotation procedures. Concretely, we first synthesize abundant training images conditioned on the semantic masks provided by realistic datasets. This yields extra well-aligned image-mask training pairs for semantic segmentation models. We surprisingly observe that, solely trained with synthetic images, we already achieve comparable performance with real ones (e.g., 48.3 vs. 48.5 mIoU on ADE20K, and 49.3 vs. 50.5 on COCO-Stuff). Then, we investigate the role of synthetic images by joint training with real images, or pre-training for real images. Meantime, we design a robust filtering principle to suppress incorrectly synthesized regions. In addition, we propose to inequally treat different semantic masks to prioritize those harder ones and sample more corresponding synthetic images for them. As a result, either jointly trained or pre-trained with our filtered and re-sampled synthesized images, segmentation models can be greatly enhanced, e.g., from 48.7 to 52.0 on ADE20K. Code is available at https://github.com/LiheYoung/FreeMask.

Paper

References (75)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer VTnr4/10 · confidence 4/52023-06-24

Summary

The authors propose to use paired synthetic data to train a semantic segmentation model. Specifically, a filtering strategy and a resampling strategy are proposed to control the quality of synthetic data. In this way, the paired synthetic data could further promote the standard segmentation model.

Strengths

- The overall idea makes sense, and the implementations are reasonable. - The paper is well written and easy to follow.

Weaknesses

- There may be conflict between filtering hard pixels and re-sampling hard masks, because the re-sampled hard samples may be filtered out. More in-depth analysis could be appended to solve this concern. - The performance gains (e.g., 48.5 → 50.6) seems not significant enough, considering that the proposed method introduce too many extra heavy processings. Specifically, 1) fine-tune the generative model on the specific dataset; 2) pre-train a segmentation model, i.e., Line 197; 3) finally train the desired segmentation model. - The effectiveness of the proposed method depends on the gap between the generative model and specific dataset, and thus the in-depth analysis about the adaption (fine-tuning) of generative model is necessary. On the one hand, the generative model is hard to adapt when the specific dataset is small. On the other hand, the generative model is harder to provide rich cases when the specific dataset is large. Some in-depth analyses could be appended.

Questions

See *Weaknesses.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

See *Weaknesses.

Reviewer jaKB7/10 · confidence 4/52023-07-04

Summary

The paper introduces an automatic dataset generation with mask-to-image translator. The proposed dataset generator enables to generate controllable and consistent semantic labels in generated images. The labels can be treated as fully supervised teachers from a generator and create pre-trained segmentation models under the (synthetic) supervised learning. Moreover, the dataset generation framework serves difficulty levels inside of the contents in an image. In the framework, the authors utilize FreestyleNet pre-trained with StableDiffusion in order to translate from mask to image (mask-to-image) for the pre-training pairs.

Strengths

- This paper is clearly written and easy to understand. The presentation to describe the proposed method (e.g., Figures 2, 3, 4) is high quality and convincing in visualization. - The paper could serve as a good example for data-driven approach in semantic segmentation tasks. - Experiments and their results cover multiple aspects of the performance of the synthetic pre-training. The paper will serve a good inspiration to others in synthetic pre-training and related topics. - The two different aspects, 'noise filtering (Section 3.2)' and 'image re-sampling (Section 3.3)' in synthetic pre-training are very reasonable approaches. Indeed, these two have been shown to be effective as shown in Tables 4 and 5. These two approaches are complementarily improve the segmentation performance in terms of mIoU.

Weaknesses

The reviewer does not find a critical weakness in the paper. However, it could be shown a little weaknesses. - The paper includes the descriptions as 'controllable' in l.38 and l.179, however, could the proposed approach make more controllable dense annotations? For example, a semantic mask can be manually edited in addition to the image dataset as is. Could the authors consider the kind of flexibly edited approach by humans? As an example of image editing and generation, the reviewer can raise GauGAN from [Park+, CVPR19]. The reviewer doesn't think any image editing should be acceptable, but it would be effective in terms of making the synthetic segmentation dataset more flexible. [Park+, CVPR19] Taesung Park et al. "Semantic Image Synthesis with Spatially-Adaptive Normalization" in CVPR 2019. - This is not a critical weakness, however, can alone with the synthetic pre-training surpass the real-image dataset? In fact, the single approach with synthetic images can reach comparative scores as shown in Tables 1 and 2. In some cases, the gap is quite close each other (e.g., 48.5 vs. 48.3 in SegFormer-B4 on ADE20k dataset). If it is possible on paper limitation, adding the discussion would make the paper more valuable.

Questions

- Can the proposed dataset generator with FreestyleNet precisely generate the training images by taking care of object boundaries? As the paper described in l.144-145, "We observe it is more precise in structures to synthesize images from masks than predict masks for images, especially in boundary regions". However, the reviewer is concerning that the quality of object boundaries may affect the recognition accuracy even if the process passed as mentioned in Section 3.2 filtering noisy synthetic regions.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

There are no negative limitations and societal impacts. On the contrary, the paper alleviate the privacy issues by means of synthetic image (pre-)training. The direction can effectively accelerate the kind of ethical problems in the future.

Reviewer FZD96/10 · confidence 5/52023-07-05

Summary

The Paper talks about using synthetic images generated using generative models as the training set to achieve stronger semantic segmentation models. The efficacy of the model is evaluated on ADE2K and COCO dataset, using SegFormer model. Authors propose pretraining with synthetic and joint images to evaluate which one works the best and in which scenario. Uses sampling and filtering training mechanism to further help the results.

Strengths

1. The paper demonstrates a clear structure, effectively explaining the fundamental modules and training procedure involved. It successfully addresses the task of scene understanding by leveraging annotations obtained from generative models, thereby enhancing the results of the semantic segmentation task. 2. The inclusion of a self-adaptive module within the training pipeline is a valuable contribution. This module effectively refines erroneous or spurious training examples, leading to improved model performance. 3. The paper incorporates a sampling strategy that focuses on hard-to-learn cases. By prioritizing these challenging instances, the overall performance of the model is enhanced, resulting in better final results.

Weaknesses

Methods used to improve the performance. 1. Mining Hard Examples: This is not something novel and (different flavors of it) has been used in many previous work to help in training better examples in case semantic segmentation and object detection. 2. Remove extraneous and harmful samples: Again the filtering mechanism proposed here is not that novel as previous works in the domain of active learning have again used similar techniques to improve training. 3. Use off-the-self models both for generation and training the model, nothing novel in that regard.

Questions

1. Would be interesting to know, what is the class frequency for the new generated dataset used to train the model. Does it follow the similar class frequency distribution as seen in ADE20K and COCO? 2. What are the classes for which the improvement is seen to be significant?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

None

Reviewer 6xi15/10 · confidence 5/52023-07-07

Summary

This paper proposes to generate densely annotated synthetic images with generative models to help supervise the learning of fully supervised semantic segmentation frameworks. To improve the effectiveness of synthetic images, the authors further design a robust filtering criterion to suppress noisy synthetic samples at the pixel and class levels and propose an effective metric to indicate the hardness of semantic masks where they sample more synthetic images for harder masks. Ablation studies validate the effectiveness of the proposed method.

Strengths

1. The logic of the article is generally clear, and the method is easy to understand. 2. The ablation experiments of this paper indicate the effectiveness of the proposed method to a certain extent.

Weaknesses

1. The proposed strategy termed re-sampling synthetic images based on mask-level hardness is somehow like “Online Hard Example Mining” (OHEM) which is widely used in computer vision area including semantic segmentation, and the adopted metric to measure the sample hardness is also the widely used average losses of all pixels in the input image. Where is the novelty of this part? I don't think generating more hard samples should be the contribution of this section since it is more likely to belong to the contribution of Section 3.1. 2. The analysis in Section 4.4 for Table 6 is not convincing in my eyes. First, after adopting “Filtering & Re-sampling”, N_max no longer denotes for the number of synthetic images used for training, thus the comparison in Table 6 is unreasonable. If I understand correctly, the number of used synthetic images should be n_p in Eq. (2), which may result in a wrong conclusion in Line 339-348. Second, where is the performance ceiling when setting N_max after adopting “Filtering & Re-sampling”? It looks like setting larger N_max (> 20) will yield better segmentation results. Finally, how to set p in Eq. (2). The reviewer does not find this detail in both paper and the supplement materials. 3. Why there are no quantitative comparisons between the proposed method and previous methods like DatasetGAN and BigDatasetGAN? The authors argue that “the main drawback of such methods is the involvement of expensive human efforts.” So, how about applying a simple pseudo labeling strategy on these generated images and comparing the results between image-pseudo labeling strategy and your proposed mask-to-image synthesis strategy? If the results are comparable, where are the advantages of using your proposed method? The reviewer believe it is important to compare your method and previous similar methods quantitatively to show the novelty of this paper. 4. It seems that the performance improvements in Table 3 for Mask2Former is limited. To my knowledge, just simply run Mask2Former for two times may also bring such improvements. Could the author give some analysis about the limited gains for Mask2Former? 5. The filtering strategy in Section3.2 is naïve and not reliable in my eyes. Specifically, could the authors provide any quantitative results to show that the proposed filter strategy could filter noisy synthetic region rather than some hard samples? 6. Previous studies like “Focal Loss for Dense Object Detection” indicate that OHEM is unreasonable. Whether the authors compare your re-sampling strategy with some objective-function-based strategy like Focal Loss, weighted cross entropy loss to show the effectiveness of your method? 7. Why higher mIoU must be a good point is Table 4? For example, if your method could generate some densely annotated synthetic images with other domains which is different from the training and testing image? Whether it would lead to the decrease of mIoU but make the segmentor adapt to broader application scenarios? The reviewer thinks the latter is more important. 8. Could the authors give the results of performing the methods in Section 3.2 and 3.3 on the real training images? I mean filter the real images with the proposed strategies and re-train the model to see the importance of the generated images.

Questions

Questions are listed in the weaknesses.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Limitations have been discussed in the paper.

Reviewer 3GaN4/10 · confidence 4/52023-07-10

Summary

This paper proposes a method of synthesizing training images and corresponding semantic masks for training a semantic segmentation network. The off-the-shelf semantic image synthesis model, FreestyleNet, is used to generate images from existing semantic masks. Following the proposed re-sampling technique based on mask-level hardness, harder samples are more frequently generated. During training, to avoid the noisy pixel hampering the model training, pixel-level ignoring technique is used. The generated synthetic images are shown to be effective for model training when they are used together with the existing fully supervised labels.

Strengths

- The paper is overall well-written and easy to understand. - It is interesting that the synthesized images can improve the performance together with fully supervised dataset. This can be practically utilized for many researchers.

Weaknesses

- The proposed method heavily depends on trained mask-to-image generative models. The authors showed that naive generation of synthetic images is not sufficient for training a segmentation network, but the filtering and re-sampling techniques are quite naive. Specifically, ignoring uncertain pixels during training is popularly used for label-efficient learning (e.g., weakly and semi-supervised semanic segmentation). - Closely related references, copy-paste methods (e.g., [ref1]), are missing. They also augment training data by synthesizing images, but unlike the proposed method, they do not require any additional heavy models. The copy-paste method should be discussed and compared. [ref1] Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation - In Abstract, the authors mentioned that "We surprisingly observe that, merely with synthetic images, we already achieve comparable performance with real ones", but I think it is overstated. To synthesize these images, the trained FreestyleNet is required, but FreestyleNet is already trained with real image-mask pairs. In addition, all the technical design (filtering, re-sampling) and values of hyper-parameters are determined with fully supervised validation data. I recommend the authors to tone down the sentence in Abstract. - I guess the global hardness in Line 229 do not consider the difficulty of segmenting small objects. Intuitively thinking, small objects of hard class should increase the global hardness, but they actually slightly contribute to the global hardness. - The authors used only the limited number of synthesized images due to the disk issue. I recommend the two additional experiments: 1) the performance change by varying the number of synthesized images. With this trend, we can infer if more synthesized images can further improve the performance. 2) Increase the total number of synthesized images by saving low-resolution images or high-tolerence polygon of masks.

Questions

- Do the authors have a plan to release the code? I strongly recommend the authors to make their code publicly available.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

No limitation is discussed.

Authorsrebuttal2023-08-14

Looking forward to further feedback

Dear Reviewer 3GaN, We are sincerely grateful to you for the precious time and selfless efforts you have devoted to reviewing our paper. We would like to inquire whether our response has addressed your concerns and if you have the time to provide further feedback on our rebuttal. We are more than willing to engage in further discussion. Best regards, Authors of paper 6586.

Authorsrebuttal2023-08-18

Further feedback

Dear Reviewer 3GaN, Thank you for your selfless efforts. As for your previous concern about discussions with Copy-Paste (our response Q1-2), in addition to our previous analysis and qualitative comparisons, we here provide more quantitative results about Copy-Paste (object-level mixing, we use ClassMix in semantic segmentation for no bounding box information) and CutMix (random rectangle-region mixing): | Real Only | Real + Copy-Paste | Real + CutMix | Real + Synthetic (Ours) | |:---------:|:-----------------:|:-------------:|:-----------------------:| | 45.8 | 45.6 (-0.2) | 45.9 (+0.1) | **48.0 (+2.2)** | In conclusion, Copy-Paste and CutMix do not help much in the *fully-supervised semantic segmentation* task. Similar observations are also reported in [1, 2] (please refer to their provided *fully-supervised* results, the *fully-supervised* results are even downgraded after applying CutMix or ClassMix). [1] Semi-supervised semantic segmentation needs strong, varied perturbations, In *BMVC*, 2020. [2] ClassMix: Segmentation-based data augmentation for semi-supervised learning, In *WACV*, 2021.

Authorsrebuttal2023-08-20

Looking forward to further feedback

Dear Reviewer 3GaN, We are sincerely grateful to you for the precious time and selfless efforts you have devoted to reviewing our paper. We have provided our detailed response to each concern. Since the deadline for reviewer-author discussion is approaching, we would like to inquire whether our response has addressed your concerns and if you have the time to provide further feedback on our rebuttal. We are more than willing to engage in further discussion. Best regards, Authors of paper 6586.

Authorsrebuttal2023-08-14

Looking forward to further feedback

Dear Reviewer 6xi1, We are sincerely grateful to you for the precious time and selfless efforts you have devoted to reviewing our paper. We would like to inquire whether our response has addressed your concerns and if you have the time to provide further feedback on our rebuttal. We are more than willing to engage in further discussion. Best regards, Authors of paper 6586.

Authorsrebuttal2023-08-14

Looking forward to further feedback

Dear Reviewer FZD9, We are sincerely grateful to you for the precious time and selfless efforts you have devoted to reviewing our paper. We would like to inquire whether our response has addressed your concerns and if you have the time to provide further feedback on our rebuttal. We are more than willing to engage in further discussion. Best regards, Authors of paper 6586.

Reviewer FZD92023-08-18

Reply to the Authors

Thank the authors for providing response in detail. I would like to agree with some of the answers that authors have responded with. About the improvement seen for the classes. The observation seems like it improves the results for some of the more less seen classes than the more often seen one at least for ADE20K dataset. My follow up questions would be 1. What could be probable reasons for improvement seen in the classes such as ship: +68.19, (2) microwave: +48.72, (3) arcade machine: +45.85, (4) booth: +45.66, (5) oven: +30.86, (6) skyscraper: +23.23, (7) swimming pool: +15.52. There does not seem to be any relation in terms of size or semantics. What are authors comment on this ? 2. Also for both COCO and ADE20K datasets, it would be interesting to see what is the improvement seen for classes which seen the most.

Authorsrebuttal2023-08-18

Further response

Dear Reviewer FZD9, thank you very much for your further feedback. For the first question about the large improvements of our listed classes, we think there are three reasons jointly contributing to this: - **[More diverse data distributions]** Our processed synthetic training pairs largely compensates the original real training dataset. For example, there are 11x more pixels belonging to *swimming pool* after introducing our synthetic set. These additional novel training samples sufficiently provide more diverse data distributions for our new training set, enabling stronger test performance and even stronger out-of-distribution (OOD) ability (as for OOD, we politely refer you to our response Q2-9 to Reviewer 6xi1). - **[Larger room to improve]** To be honest, there is much larger room to improve on these classes. For example, the baseline IoU (trained merely with real images) is only 10.82% on the *ship* class, 36.19% on the *arcade machine* class, and 30.06% on the *booth* classes. - **[Promising synthesis quality]** These classes are indeed not rare in the training set of our pre-trained Stable Diffusion (SD) model. Thus, the synthesis quality of these classes is acceptable. For example, according to our search, there are 15,473 *ship*-relevant images and 5,431 *swimming-pool*-relevant images on the SD training set LAION-5B. After further fine-tuned on our target set (*e.g.*, ADE20K), the synthesis ability can be further enhanced. For the second question, the most frequent class on ADE20K is *wall*, and we improve its IoU from 77.51% to 78.65%. The most frequent class on COCO is *person*, and we improve its IoU from 85.15% to 85.95%. These two improvements are already significant, because the real training samples are already abundant on the two classes and the baseline results are rather competitive (classes with nearly 80% IoU are very challenging to improve under the same model architecture, like the Pascal VOC and Cityscapes benchmarks). We sincerely appreciate your precious time and further feedback. We are wondering whether our further responses have fully addressed your concerns. Thank you very much.

Reviewer FZD92023-08-18

Reply to Authors

Thank you for providing further details about the quantitative results on class specific results. It is good to know that "there are 11x more pixels belonging to swimming pool after introducing our synthetic set. " It would be good to add these statistics for most improved classes for both ADE2K and COCO dataset in the final version. Giving us a sense about how exactly does this impact the final performance. Given the final response I would change my rating to Borderline Accept.

Authorsrebuttal2023-08-18

Thank you for your appreciation of our response

Dear Reviewer FZD9, thank you for acknowledging our response and appreciating our work. We will follow your advice and draw a detailed bar chart to demonstrate the class-wise increased sample number and class-wise IoU improvement in our final version. Thank you for being willing to change your rating to Borderline Accept. But it seems the score has not been modified. Would there be any further concerns we can help?

Area Chair jXMB2023-08-21

Dear authors, with FZD9's raise to weak accept I consider the concerns addressed and your question answered.

Authorsrebuttal2023-08-14

Looking forward to further feedback

Dear Reviewer VTnr, We are sincerely grateful to you for the precious time and selfless efforts you have devoted to reviewing our paper. We would like to inquire whether our response has addressed your concerns and if you have the time to provide further feedback on our rebuttal. We are more than willing to engage in further discussion. Best regards, Authors of paper 6586.

Authorsrebuttal2023-08-17

Looking forward to further discussions

Dear Reviewers 3GaN, FZD9, and VTnr, We are deeply grateful for your precious time and constructive feedback. For all raised concerns, we have provided necessary quantitative results or detailed explanations. We are greatly motivated that Reviewer 6xi1 has improved the score from 3 to 5 (borderline accept) and Reviewer jaKB gives a score of 7 (accept). We are wondering if there are any remaining concerns from you. We are more than willing to provide further clarifications on any aspect. Thank you very much. Best regards, Authors of paper 6586.

Authorsrebuttal2023-08-20

Looking forward to further feedback

Dear Reviewer VTnr, We are sincerely grateful to you for the precious time and selfless efforts you have devoted to reviewing our paper. We have provided our detailed response to each concern. Since the deadline for reviewer-author discussion is approaching, we would like to inquire whether our response has addressed your concerns and if you have the time to provide further feedback on our rebuttal. We are more than willing to engage in further discussion. Best regards, Authors of paper 6586.

Authorsrebuttal2023-08-21

Dear Reviewer VTnr, thank you very much for your precious time and great contributions. Since we are approaching the discussion deadline in 10 hours, we would like to ask if there is any further feedback on our rebuttal. We are greatly motivated that Reviewer 6xi1 has improved the score from 3 to 5 (borderline accept), Reviewer FZD9 has improved from 4 to 6 (weak accept), and Reviewer jaKB keeps a score of 7 (accept). Besides, we believe our work has improved a lot from your constructive feedback. For example, we will add detailed experiments and discussions about the effectiveness of our method in case of small or large datasets (also provided above in the [rebuttal](https://openreview.net/forum?id=XOotfgPiUF&noteId=d5NhljZDiI)). We are eagerly looking forward to your further attention. Thank you very much. Best regards, Authors of paper 6586.

Area Chair jXMB2023-08-21

Please provide feedback to authors

Dear Reviewer VTnr, what are you concerns after the rebuttal and in light of the other reviews? Could you please provide some feedback to the authors? Thanks,

Reviewer 3GaN2023-08-20

I appreciate the authors' response. My minor concerns are addressed, but I still think the strength of the proposed method comes largely from the superiority of FreestyleNet. The filtering and hard negative sampling seems trivial to me. I would keep my original rating

Authorsrebuttal2023-08-20

Dear Reviewer 3GaN, thank you very much for your further feedback. It is glad to know part of the concerns have been addressed. As for the remaining concern about whether the strength of our method comes largely from FreestyleNet instead of our proposed two processing techniques, we would like to provide some further clarifications here. - First, directly using the synthetic pairs from FreestyleNet *without any processing* **will not bring any improvement** to the challenging fully-supervised real-image baseline. This has been supported by Table 5 in our main paper. We borrow the results below (the real-only result is borrowed from main Table 3 with SegFormer-B4). The inferior result of blindly integrating synthetic images (FreestyleNet, -0.2 mIoU) also explains why existing works [1, 2] mostly only address the few-shot scenario, which contains a few real images and is easy to boost. In comparison, with our filtering and hardness-aware synthesis strategies, we can significantly boost the challenging but practical fully-supervised scenario (**+1.8 mIoU**). | | Real data | Synthetic data *w/o processing* | Synthetic data *w/ processing* | validation mIoU | :-: | :-: | :-:| :-:| :-: | Real only | ✔ | | | 48.5 | | FreestyleNet | ✔ | ✔ | | 48.2 (-0.3) | | **Ours** | ✔ | | ✔ | **50.3 (+1.8)** | - Second, we have further validated the necessity and superiority of our two strategies in Table 6 of our main paper. Results are borrowed below. We observe that, *without our proposed techniques*, scaling up the number of synthetic images *does not* yield consistent improvement, even gradually downgrading the performance. In addition, our proposed method tremendously outperforms the plain FreestyleNet which does not include any processing techniques. The performance gap can be as large as **+5.0 mIoU**. | $N_{\max}$ (scaling ratio) | 6 | 10 | 20 | :-: | :-: | :-:| :-: | w/o processing (FreestyleNet) | 43.7 | 43.6 | 43.3 | | **w/ processing (Ours)** | 47.2 | 47.7 | **48.3** | | Superiority over FreestyleNet | +3.5 | +4.1 | **+5.0** | - Lastly, as we provided in the [rebuttal](https://openreview.net/forum?id=XOotfgPiUF&noteId=k6Zw1ZF8r9), our hardness-aware synthesis strategy is not trivial. It is much better than online hard example mining (OHEM). We also borrow the comparisons below (44.2 *vs.* 45.4). As for your opinion that similar filtering strategies also exist in label-efficient learning, we agree with this. Indeed, this is a very common motivation. Similar practices also exist in noisy label learning. However, as we clarified in the global response, *we use similar motivations for totally different purposes*. We aim to recognize the synthesis failure cases, which is proved fundamental to the success of later perception task. The necessity of "cleaning" synthetic images is almost completely ignored in previous works [3] that utilize synthetic images to benefit perception tasks. *We believe that, drawing attention to "more effectively" learning from synthetic images, rather than always focusing on better synthesis quality, is also our contribution to the community.* |Baseline|+ OHEM (thresh=0.7, min kept=100K)|+ Our hardness-aware synthesis |:-:|:-:|:-:| |44.0|44.2|**45.4**| Please tell us if you have any further concerns about this response. We are more than willing to provide any further explanations. Thank you very much for your precious time and great contributions. [1] He, Ruifei, et al. "Is synthetic data from generative models ready for image recognition?." *ICLR* 2023. [2] Wu, Weijia, et al. "Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models." *ICCV* 2023. [3] Azizi, Shekoofeh, et al. "Synthetic data from diffusion models improves imagenet classification." *ICCV* 2023.

Authorsrebuttal2023-08-21

Dear Reviewer 3GaN, thank you very much for your previous feedback. Since we are approaching the discussion deadline in 10 hours, we would like to ask whether there are any further concerns about our work. There is only one remaining concern about whether the strength of our method comes largely from FreestyleNet. We have provided comprehensive experiments above (borrowed from our main paper) to prove that without our processing techniques, the synthetic pairs from FreestyleNet cannot benefit the real-image baseline, even downgrading the performance. We believe our results are convincing. We are eager for your further attention. We appreciate your precious time a lot. Best regards, Authors of paper 6586.

Reviewer jaKB2023-08-21

Response to author rebuttal

Thank you so much for the authors. The Q&A is one of the most important things in the future generative AI and learning. The reviewer encourage the authors, and will keep the paper rating as "7: Accept". Thanks again for the discussion.

Authorsrebuttal2023-08-21

Dear Reviewer jaKB, We are deeply grateful for your great appreciation of our work. Thank you so much for all the precious time you have devoted and the constructive feedback you have provided. We will include all your feedback in our final version. Best regards, Authors of paper 6586.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC