Thank you so much for the positive rating and insightful comments. Your valuable suggestions are very helpful for further strengthening our paper.
**[1. The three components have been proposed in the related work]**
The core contribution of our paper lies in the introduction of a new paradigm – "dense prediction, followed by sparse refinement" – which paves the way for efficient, high-resolution semantic segmentation. We incorporate existing modules to provide one possible instantiation of the new paradigm, but this should not be mistaken as a lack of novelty. Our focus isn't on the design of any specific module, but rather on the novel sparse refinement paradigm itself. Each component can be readily substituted to create a completely new implementation.
**[2. Proper validation of hyper-parameter $\alpha$]**
We ensured that the test data was not utilized during the selection process of α. Our guiding principle for α selection is based on the desired recall ratio. We consistently aim for a recall ratio of approximately 80 percent, which is sufficient to achieve comparable results. Of course, a higher recall ratio can yield improved outcomes, but it also leads to increased running time.
**[3.Training time]**
You may misread the paragraph. The training time is 96 RTX A6000 hours, not days.
**[4. Capacity of the sparse feature extractor]**
We experimented with different model capacities and the results are shown in the following table. Specifically, we incrementally increase the capacity of MinkUNet by expanding channels and adding more stages. The results indicate that performance improves as the model becomes larger, but eventually reaches saturation at 80.9 mIoU.
| | #of Channels(32$\times$) | mIoU
| ----------- | ----------- | ----- |
| MinkUNet | {1,2,4,8} | 80.5 |
| MinkUNet | {1,2,4,8,16} | 80.9 |
| MinkUNet | {1,2,4,8,12,16,24,32} | 80.9 |
**[5. Was MinkowskiUNet pre-trained or trained from scratch]**
We train it from scratch.
**[6. Show the accuracy for 100% density in Table 5a]**
The result is 80.5 when incorporating all the pixels, lower than 80.9 we got with 12% density. This is reasonable because our goal is to refine the pixels that are "difficult to learn in the low-resolution" ones. However, incorporating all the pixels would also involve including many pixels that are easy to learn. This can serve as a shortcut for the model to easily achieve low loss with weaker capability.
**[7. Include simple average ensembling]**
The result is 80.4 when using simple average ensembling, showing the effectiveness of our ensembler.
**[8. Explain the magnitude selector.]**
The magnitude selector calculates the L2 magnitude on the output of the last layer, just before the segmentation head, in order to obtain an importance score for each pixel. This approach is commonly employed in token pruning works to identify and remove unimportant areas. However, the unsatisfactory results obtained from the magnitude selector highlight the lack of a strong correlation between importance and uncertainty.
**[9. Show MACs in Table 7]**
The #MACs for the cuBLAS backend remains the same as the high-resolution model due to the dense input. And the #MACs for SpConv and TorchSparse are the same, shown in the main table. For instance, in the case of HRNet-W48 on Cityscapes dataset, the cuBLAS backend has #MACs of 0.75T, while SpConv and TorchSparse have #MACs of 0.32T.
**[10. Report minimal hardware requirements]**
The GPU RAM required for the task is approximately 20G per GPU, making it compatible with GPUs such as the GeForce RTX 3090 equipped with 24G RAM.
**[11. Consider citing earlier work]**
Thanks! We will cite these works in our revision.
**We hope our response has resolved all of your concerns. Please do not hesitate to contact us if there are other clarifications or experiments we can offer. Thank you!**