# review: Open-Source Data-Driven Cross-Domain Road Detection from Very High Resolution Remote Sensing Imagery

A. Suitability of Topic
1. Is the topic appropriate for publication in these transactions?

Adequate Match

B. Content
1. Is the paper technically sound?

Yes

2. How would you rate the technical novelty of the paper?

Not Novel

Explain:

The paper proposed a framework for domain adaptation of for road detection and also use the open-sourced OSM data to generate road segmentation labels for the target domain.
The domain adaptation module is a combination of existing techniques including cycle consistency [R1] and perceptual loss [R2], although the performance is somewhat better than other tested method, the reason why the proposed method is effective is not very well discussed, and given that cycle consistency and perceptual loss are well known so the novelty of the proposed domain adaptation module is limited.
On the asepct of using the OSM data to generate segmentation labels on the target domain, the technique of using OSM data to generate label for training a neural network is also not novel, this has explored by [R3] and [R4], and the paper does not provide any in depth investigation and explanation of how does the label generated from OSM data helps the learning of the neural network.

[R1] CyCADA: Cycle-Consistent Adversarial Domain Adaptation
[R2] Perceptual Losses for Real-Time Style Transfer and Super-Resolution
[R3] Road Extraction from Very High Resolution Images Using Weakly labeled OpenStreetMap Centerline
[R4] Building High Resolution Maps for Humanitarian Aid and Development with Weakly- and Semi-Supervised Learning

3. Is the contribution significant?

Incremental

4. Is the coverage of the topic sufficiently comprehensive and balanced?

Treatment somewhat unbalanced, but not seriously so

5. Rate the Bibliography

Unsatisfactory

C. Presentation

1. How would you rate the overall organization of the paper?

Could be improved

2. Are the title and abstract satisfactory?

No

3. Is the length of the paper appropriate? If not, recommend what should be added or eliminated.

No (Explain):
In the experiments section, I would expect the paper to include more ablation studies to show how effective of the label generated from OSM data since the paper includes it as a contribution.
For example, 1) use a different length for the buffer to generate the mask instead of 3.5m. 2) generate more data with labels than 3000 samples to test the extend of the benefit from OSM data.
For the proposed domain adaptation module, multiple components are proposed for training with multiple losses, however, the experiments does not show any ablations on that, for example, how will the final performance be influenced when the user remove one of the losses or components?

4. Are symbols, terms, and concepts adequately defined?

Not always

5. How do you rate the English usage?

Needs improvement

Recommendation

R - Reject

Comments to the Author

1. I'd suggest the author to perform a more detailed ablation of the proposed method, speifically about the labels generated from OSM, it seems that using 3000 samples from OSM alone can already achieve a good performance on Birmingham images, (F1=0.7515 for OSM alone, and 0.7618 for the full methods), can using more samples generated from OSM further improve the performance or even surpasses the performance of the proposed domain adaptation module?
2. The proposed domain adaptation module seems to be not very effective considering the number of components during training and the final performance, from Table V and IV, I think the most performance gain is from the OSM data, and the proposed domain adaptation module alone does not seem to be effective and even underperforms previous methods.
3. The domain adaptation module is too complexed, and the description in section III A are not very clear to understand, the G_{st}^{s} and G_{st}^t in line 24 page 3 are not defined when they first appear. and also in eq (1), the loss L_{seg_{adv}} should be L_{seg\_adv}.
4. The comparison with other methods are not comprehensive, there exists many other methods for domain adaptation for segmentation other than just a global adversarial network such as [R1,R2,R3], just to name a few.

[R1] Generalize Then Adapt: Source-Free Domain Adaptive Semantic Segmentation. ICCV2021.
[R2] Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation. CVPR2021.
[R3] Instance Adaptive Self-Training for Unsupervised Domain Adaptation. ECCV2020.