Monocular depth estimation (MDE) is a fundamental topic of geometric computer vision and a core technique for many downstream applications. Recently, several methods reframe the MDE as a classification-regression problem where a linear combination of probabilistic distribution and bin centers is used to predict depth. In this paper, we propose a novel concept of iterative elastic bins (IEBins) for the classification-regression-based MDE. The proposed IEBins aims to search for high-quality depth by progressively optimizing the search range, which involves multiple stages and each stage performs a finer-grained depth search in the target bin on top of its previous stage. To alleviate the possible error accumulation during the iterative process, we utilize a novel elastic target bin to replace the original target bin, the width of which is adjusted elastically based on the depth uncertainty. Furthermore, we develop a dedicated framework composed of a feature extractor and an iterative optimizer that has powerful temporal context modeling capabilities benefiting from the GRU-based architecture. Extensive experiments on the KITTI, NYU-Depth-v2 and SUN RGB-D datasets demonstrate that the proposed method surpasses prior state-of-the-art competitors. The source code is publicly available at https://github.com/ShuweiShao/IEBins.
Paper
Similar papers
Peer review
Summary
The paper introduces an Iterative Elastic Bins (IEBins) approach for monocular depth estimation. Many conventional monocular depth estimation approach uses a soft-argmax representation (Eq. (4)) that sums the product between depth probability and pre-defined depth bins. However, the large number of pre-defined bins hinders model convergence (@L.43). To alleviate the problem, the paper proposes to use adaptive ranges of depth bins for each pixel based on previous intermediate depth estimates. Along with the number of iteration steps, the depth ranges get smaller for fine-detail estimation (yet, taking uncertainty into account). The paper shows empirically good accuracy over previous methods.
Strengths
+ Good results Table 1, 2, 3, and 4 show that the method achieves better accuracy than published methods in the public benchmark datasets (KITTI Eigen, KITTI, NYU-Depth-v2, and SUN RGB-D). + Good ablation study Table 5 and 6 demonstrates how each design choice affects/improves the accuracy. Table 5 shows the strength of the proposed IEBin idea over previous approaches (UBins, SIBins, AdaBins, and IBins). Table 6 also justifies the design choice of the number of bins.
Weaknesses
- Question on the number of stages (Fig. 3) The method uses 6 stages (iteration steps) in total. I wonder what happens if it uses more iteration steps during the inference (i.e., 6 iterations during training, but 8 or 10 iterations during test time). Will the method keep improving the accuracy with more iteration steps? Or, even during training, what if it uses more iteration steps other than 6? How the number 6 is set? - Possibly an unfair comparison? Table 5, Comparison with different bin types. I wonder if other methods (UBins, SIBins, AdaBins, and IBins) also use the iterative optimizer. For a fair comparison, I think it's also good to prepare a baseline with the iterative optimizer and only change the depth bin types. Then it can differentiate the performance gain of bin types from the iterative optimizer. Despite the good results, I would like the paper Borderline Reject for now, mainly due to the possible unfair comparison in Table 5. The source of gain is not so clear if it's from the iterative optimizer or the depth bin types. ---- I share the same concern with Reviewer 4WPj that the main contributions don't seem to be so strong. The iterative refinement ideas have been proposed in other literature (e.g., optical flow, SfM), and the paper demonstrates empirical accuracy gain. That said, I am raising the score to Borderline Accept as the authors' response resolves my main concerns. I hope all the comments and concerns during the discussion phase are reflected in the updated version.
Questions
- In Fig. 2 the three depth maps on the right look the same. Probably it may need to be changed to the actual experiment results. - In the last example (chair) in Fig. 5, the object boundary near the chair is still blurry. (Just out of curiosity) I wonder if the proposed representation can resolve this blurry object boundary problem because it limits the depth range, so I would expect it outputs much clearer depth around object boundaries. - In Table 7, what is the source of the faster inference time than BinsFormer?
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited 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
The paper didn't include its limitations or societal impact.
Summary
The paper introduces a method for monocular depth. The method uses a recurrent network based on RAFT to predict a probability distribution over a set of bins, which enables the depth at each iteration to be computed as the expectation over all bins and the margins of each bin to be adjusted using the computed variance. In the subsequent iterations, a finer-grained search is performed in the adjusted bins. This new strategy is introduced as "IEBins". The authors motivate this strategy as being robust to uncertain initial predictions.
Strengths
The method outperforms prior work across multiple standard evaluation settings. The method is interpretable, producing confidence for different regions and steadily improving depth predictions. To the best of my knowledge, the iterative design is substantially different than the usual approach for monocular depth. The introduced method is very fast. The IEBins are compared against other binning strategies in ablation experiments.
Weaknesses
No obvious weaknesses.
Questions
None
Rating
8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.
Confidence
2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
Soundness
4 excellent
Presentation
4 excellent
Contribution
4 excellent
Limitations
I did not see discussion of limitations
Summary
This paper introduces a classification-regression-based monocular depth estimation pipeline. Previous classification or classification-regression-based MDE approaches often suffer from huge complexity and loss of generalizability issue, due to the nature of requiring more depth hypothesis for better performance. In this paper, it propose iterative elastic bins, ie, IEBins technique, which iteratively adjusts bin ranges as the prediction converges, so that with less number of bins and fast inference time, the method can still perform on the state-of-the-art level. It additionally introduces dedicated transformer-based feature extractor and GRU-based iterative optimizer.
Strengths
This paper is well-written and easy to follow. The proposed IEBins is an intuitive and interesting idea, and the performance also backs up its effectiveness.
Weaknesses
More in-depth analysis on 'iterative' perspective of IEBins will be appreciated. 1) For example, what happens when iteration is less or more than 6? How much does the number of iterations affect the final performance? 2) The authors claim that the proposed GRU-based iterative optimizer is also one of their contributions, saying that it is helpful to capture temporal information during IEBins-based depth estimation. Yet, there is no ablation study regarding this. Comparison between adding/excluding GRU unit, alongside its effectiveness regarding the number of iteration will help strengthening the claimed contribution.
Questions
1. I have some questions regarding updating rules of the proposed IEBins. 1-1. In Eq. 7, it denotes that elastic target bin edges are modified using uncertainty from previous stage's probability distribution. Does it mean that for each pixel, bin edges are set differently? If this is the case, the network should remember separate bin ranges for every pixel, which seems like to require huge memory. Correct me if I'm wrong. 1-2. In L133, the paper says that d_min and d_max values are updated with new bin edges. In the next iteration, are 16 bins additionally set within this new min-max range? Then, does it mean that in 2nd iteration, for example, new bin range are approximately 1/256 from original min-max range? 2. In Fig 5, the range of the proposed method seems significantly off compared to GT. How is the visualization done? 3. In Tab 6, why does the performance drop when using 32 bins instead of 16? Is this aligned with overfitting issue stated in the introduction section? If it's true, than doesn't it mean that the proposed method already suffers from 32 bins while AdaBins, for example, can operate until 256 bins? How can this phenomenon be analyzed?
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
2 fair
Presentation
2 fair
Contribution
2 fair
Limitations
The authors did not address any limitations. Yet, other than weakness and questions I commented above, I don't see any severe limitation to this paper.
Summary
The paper tackles the task of monocular depth estimation which is of fundamental importance in computer vision and has many downstream applications. Several recent works use bin-based approaches and follow the adaptive binning framework where the distribution of bin centers on the depth interval (that are treated as depth candidates) can vary per image or per-pixel. This paper proposes a novel approach to the adaptive binning framework called IEBins. Instead of working with only one (potentially large) set of bins (per image or per pixel), IEBins refines the binning structure in an iterative manner. Starting with a coarse uniform division of the original depth interval (dmin, dmax), the idea is to recursively find and uniformly divide the ‘target’ bin by using the ’target’ bin as the new target depth interval for the next step. In addition, the target bin is made elastic (the new target depth interval’s ends are allowed to change) based on the uncertainty estimate. The work achieves state-of-the-art results on popular benchmarks including KITTI and NYU-Depth-v2 and the authors promise to release the code and models publicly.
Strengths
* State-of-the-art results. The work achieves SOTA in a highly competitive space of monocular depth estimation with over 7% improvement in RMSE over the prior SOTA. IEBins idea is also validated and shows about 2% improvement to adaptive bins baseline in fair settings. * Interesting idea. The idea of recursive division of bins is obvious in retrospect yet creative. Elastic nature of bins based on uncertainty also makes sense and subtly introduces the ‘adaptive’ nature of bins. * Well exploited ideas and good architecture design. Authors have well exploited the ideas from other works to obtain their goal. e.g. Iterative refinement using GRU, developed in separate field of optical flow estimation, has been introduced here in a well designed architecture. * Potential to be foundational. Recently, depth estimation has seen foundational ideas like adaptive binning that serve as a starting point for a series of several follow up works. I believe the iterative refinement of bins can prove to be as influential. * Good to see ablations on all bin types in fair settings.
Weaknesses
* W1 **Incomplete literature review**. Authors have missed some important published works that are highly relevant to this work: * [a] - LocalBins (ECCV’22) also introduces the idea of “splitting” the bins and step-wise refinement of bins in a coarse-to-fine manner. IEBins and LocalBins are highly related and pursue the same goal but follow different approaches. Quantitative comparison with [a] and providing the insights to differences and similarities is highly suggested. * [b] - PixelFormer (WACV’23) is also based on the adaptive binning framework and introduces layer-wise refinement of “pixel queries”. * W2 **Absence of quantitative evidence towards the working of the fundamental idea**. According to the proposed idea of iteratively making the ‘target’ depth interval smaller, the bin-width (or its median across an image) should exponentially decrease stage-by-stage, with some flexibility introduced by the elastic nature. At the same time, if the bins are too elastic i.e. final bin-widths are comparable to initial bin-widths (e.g. only 50% smaller than previous or so), then the iterative restrictive nature of depth search is invalidated. Although authors visualise the refinement process in Fig. 3 but these results can still be explained by ‘too elastic’ bins as the elastic-bin-adjustment varies spatially. Also, are the uncertainty 'heat maps' scaled individually or globally? If individually (which seems to be the case), then the uncertainty visualizations only show that uncertainty tends to take much higher values near edges, rather than an overall absolute decrease and therefore tell nothing about the bin widths. The authors are suggested to provide evidence to show that method actually works as described to avoid misleading conclusions. For example, via a simple line plot of evolution of target bin-width (multiple lines for pixels across a row or single line plot of median elastic bin width across image etc) through the course of refinement, or any other form that authors deem suitable that delivers the evidence clearly. * W3 **Unsubstantiated claim at L44**. Authors claim that large number of bins are undesirable and use this as the main motivation for the work. Authors state at L43-44 that “Learning ambiguity grows rapidly as the number of bins increases” but fail to provide any explanation, reference or evidence. As this information can be potentially misleading, authors should either remove the claims or provide an explanation, reference or evidence towards the claim. [a] Bhat, Shariq Farooq, Ibraheem Alhashim, and Peter Wonka. "Localbins: Improving depth estimation by learning local distributions." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022. [b] Agarwal, Ashutosh, and Chetan Arora. "Attention attention everywhere: Monocular depth prediction with skip attention." Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2023.
Questions
* How do the bin-widths actually evolve during the iterative refinement? What is the mean of elasticity factor (new adjusted width divided by the bin-width at that stage with no elasticity)? * What is the maximum bin-width in meters reached at the final stage e.g. for the NYU test set? This can also answer the elasticity to some extent. * Why is a large number of bins undesirable? What is the evidence? * How does IEBins compare with LocalBins?
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
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
2 fair
Contribution
4 excellent
Limitations
The authors do not explicitly reflect upon the limitations or scope of the work. Refer to "weaknesses" for major concerns.
Summary
This paper proposes iterative elastic bins (IEBins), a multi-stage coarse-to-fine method for monocular depth estimation (MDE). It progressively searches the target depth bin on top of the previous step. To reduce the error accumulation in the iterations, an elastic bin is proposed whose width is adjusted based on the depth prediction uncertainty. Experiments on the KITTI, NYU-v2, and SUN RGB-D datasets show improved performances on MDE.
Strengths
- The depth bin is adjusted elastically based on the depth prediction uncertainty.
Weaknesses
- The contributions are not enough to my knowledge. I think only the elastic bins are new. The iterative manner for depth estimation by GRUs has been proposed in RAFT[28] and adopted in many follow-up works, e.g., RAFT-Stereo[29] and Itermvs[30]. The depth bins as shown in Eq 1,2,3&4 have been investigated in Adabins [3]. As for the framework of the feature extractor and GRU-based layers, standard implementations are involved. The multi-stage depth estimation pattern was also seen in previous works, e.g., CasMVS [Gu et al, cvpr2020] and IterMVS[30]. - Some minor ones: - line 141: a RGB --> an RGB - line 151: during iteration --> during iterations
Questions
- Have other depth uncertainty presentations been investigated or compared with the variance of the probabilistic distribution used in Eq 6?
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
2 fair
Presentation
2 fair
Contribution
2 fair
Limitations
- N/A for the limitations. - I suggest the authors provide the failure cases (and the corresponding explanations) of the proposed IEBins, w.r.t. for example, the number of bins and the elastic width of bins.
Summary
The paper proposes an iterative elastic bins (IEBins) strategy for monocular depth estimation. The IEBins use a small number of bins adaptively at each iteration. It use a GRU to predict the depth distribution at each stage. The authors conduct experiments on 3 commonly used datasets and the proposed method shows better results than previous methods.
Strengths
The paper is well written and easy to follow. The idea of using iterative update is interesting. The experiments and ablation study are thorough and validate the proposed method's effectiveness.
Weaknesses
1. The motivation of iterative updates is not very clear. RAFT used iterative updates because at each iteration, a new cost volume can be constructed based on current prediction. But I don't see where the new information comes from in MDE. 2. Based on 1. Maybe the improved performance come from the depth candidates projects since the model knows what the depth is. Would be interesting to see how the model performance by using a positional encoding of depth candidates to other non-iterative methods, such as adabins.
Questions
L44. "This learning ambiguity grows rapidly as the number of bins increases, exacerbating the difficulty of model convergence and prone to overfitting." Why the ambiguity grows in this case? And why the proposed iterative method doesn't lead to increased ambiguity?
Rating
5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited 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
3 good
Limitations
1. Currently only the depth maps are shown. It would be great to compare the point cloud with other methods since it's much easier to see the 3D structure in point cloud. 2. How is the improved depth estimation benefits in downstream task (compared to other MDE methods)? For example, could it be used for SLAM or 3D reconstruction from multiview images.
Discussion
Thanks for sharing your responses! However, there are some unclear thoughts after reading the response. > W1: Question on the number of stages (Fig. 3) What does it mean by that ``` When the number of stages exceeds 6, the performance changes very little.```? Does it mean that the accuracy has plateaued? It would have been much great if the paper provided an analysis on trying out different numbers of iteration steps during both training and testing (even trying out different training/testing numbers, e.g., training 6 iterations, testing 8 iterations). > Q2: In the last example (chair) in Fig. 5, the object boundary near the chair is still blurry. I am sorry but I didn't understand the answer clearly. Could you elaborate more on why ```the proposed method is sensitive to the object boundaries due to large depth variations in those regions```?
Discussion
### __Thank you very much for your feedback.__ ### _W1: What does it mean by that when the number of stages exceeds 6, the performance changes very little.? Does it mean that the accuracy has plateaued? It would have been much great if the paper provided an analysis on trying out different numbers of iteration steps during both training and testing (even trying out different training/testing numbers, e.g., training 6 iterations, testing 8 iterations)._ A1: As suggested, we have provided the results of different stages below (the number of stages for training and inference remains the same). |Stage Abs Rel RMSE log10 $\delta <1.25$ $\delta <1.25^2$ $\delta <1.25^3$|| |:---|:---| |Stage1 0.093    0.333   0.041    0.921     0.991       0.998| |Stage2 0.090    0.325   0.040    0.927     0.991       0.998| |Stage3 0.089    0.320   0.039    0.931     0.991       0.998| |Stage4 0.088    0.317   0.038    0.933     0.992       0.998| |Stage5 0.087    0.315   0.038    0.935     0.992       0.998| |Stage6 0.087    0.314   0.038    0.936     0.992       0.998| |Stage7 0.087    0.313   0.038    0.935     0.992       0.998| As the stage increases, the performance gradually improves until saturated, and when the number of iterations exceeds 6, the performance changes very little. Then we try to train using 6 stages and use more stages in inference phase. The results are as follows: |Stage Abs Rel RMSE log10 $\delta <1.25$ $\delta <1.25^2$ $\delta <1.25^3$|| |:---|:---| |Stage6 0.087    0.314   0.038    0.936     0.992       0.998| |Stage7 0.087    0.314   0.038    0.936     0.992       0.998| |Stage8 0.087    0.314   0.038    0.936     0.992       0.998| As we can see, the performance becomes plateaued when using 7 stages and 8 stages at inference phase in this case. ### _Q2: Could you elaborate more on why the proposed method is sensitive to the object boundaries due to large depth variations in those regions?_ A2: Generally, the depth variations are large at object boundaries. In these regions, due to the large depth variations, the proposed method can easily classify respective region into different depth ranges at initialization stage and further iteratively refine these depth ranges in subsequent stages. Hence, it is possible for the proposed representation to solve this blurry object boundary problem.
Thanks for providing the detailed analysis! It resolved my concern. > Q2: Could you elaborate more on why the proposed method is sensitive to the object boundaries due to large depth variations in those regions? I see. My original question is that why the proposed method still give blurry object boundaries on the chair example (Fig. 5, the bottom row) although it's expected to solve this problem via the iterative refinement? In Fig. 5, the proposed method gives sharper estimation than at least other methods, but (in my humble, subjective opinion) the depth boundaries still look blurry. I am not trying to pick on this single example, but trying to understand the upper bound performance of this method -- what the method can solve and cannot solve. Any discussion on this would be appreciated!
### __Thank you very much for your prompt feedback.__ ### _Q2: My original question is that why the proposed method still give blurry object boundaries on the chair example (Fig. 5, the bottom row) although it's expected to solve this problem via the iterative refinement? In Fig. 5, the proposed method gives sharper estimation than at least other methods, but (in my humble, subjective opinion) the depth boundaries still look blurry. I am not trying to pick on this single example, but trying to understand the upper bound performance of this method -- what the method can solve and cannot solve. Any discussion on this would be appreciated!_ A1: Nice comments. We think there are two possible reasons: First, we use the classification-regression to estimate depth. Compared with the classification, the depth acquired by classification-regression is smoother and more continuous, but at the same time it may blur the boundaries due to the weighted average of all depth candidates. Second, we use a pixel-wise loss to supervise the depth without imposing a direct supervision signal on the probabilistic distribution. This does not guarantee that the depth candidate corresponding to the peak point of the probability distribution is the true optimal depth candidate, thereby affecting the depth estimate.
Thanks for the comments! I hope this discussion will be included in the main paper or limitation section.
### __Thank you for the feedback.__ We will include the above discussion in the main paper or limitation section as suggested.
I thank the authors for their rebuttal. W2. The performance boost obtained from the GRU-based optimizer is around 0.001-0.005, RMSE-wise, which is not significantly high. Also, it degrades results on earlier stages of iterations. Can authors provide more insight or arguments regarding this? Q2. Minor note. I think it would be more effective to set the min-max normalizing values same throught a single sample. From what is given now, it is hard to know if one only preserves fine-grained detail or it actually predicts precise relative depth map.
Discussion
### __Thank you very much for your feedback.__ ### _W2: The performance boost obtained from the GRU-based optimizer is around 0.001-0.005, RMSE-wise, which is not significantly high. Also, it degrades results on earlier stages of iterations. Can authors provide more insight or arguments regarding this?_ ### _W2.1: The performance boost obtained from the GRU-based optimizer is around 0.001-0.005, RMSE-wise, which is not significantly high._ A1: As we can see from the results, the advantage of the GRU-based iterative optimizer is not obvious when the number of stages is 1, which may be due to the fact that there is no historical hidden state available. As the number of stages increases, the advantage of the GRU-based iterative optimizer gradually increases and tends to be stable. The GRU unit we use in our iterative optimizer is very lightweight and only contains three separable 5x5 convolution kernels. As for the number of parameters introduced, the performance gains brought by the GRU unit is competitive. ### _W2.2: Also, it degrades results on earlier stages of iterations. Can authors provide more insight or arguments regarding this?_ A2: When the stage is 7, the depth candidates are highly close. In this case, it is very difficult for the iterative optimizer to classify the true optimal depth candidate and it is easy to make mistakes, which may cause slight performance degradation. ### _Q2: Minor note. I think it would be more effective to set the min-max normalizing values same throught a single sample. From what is given now, it is hard to know if one only preserves fine-grained detail or it actually predicts precise relative depth map._ A3: We will revise our paper according to this nice advice.
I thank the authors for their response. While I think the IEBins design itself is noteworthy, I'm not totally convinced by additional contributions (such as GRU-based optimizer) and analysis. Thus, I keep my score as borderline reject.
Thanks for your reply. I think the MDE metrics are kind of saturated and would be great to show the benefits of improvements in downstream task such as slam or 3D reconstruction.
### __Thank you very much for your feedback.__ ### _L2: I think the MDE metrics are kind of saturated and would be great to show the benefits of improvements in downstream task such as slam or 3D reconstruction._ A1: As suggested, we integrate IEBins and NeWCRFs [6] into ORB-SLAM2 [1] in the RGB-D setting and evaluate the visual odometry performance on the KITTI odometry dataset. We report results on keyframes (selected by the ORB-SLAM2) and on all frames of sequences 01-10. The ATE (m) metric is used. ''key'' and ''all'' stand for keyframes and all frames, respectively. |Seq IEBins (key) NeWCRFs (key) IEBins (all) NeWCRFs (all)|| |:---|:---| |01      117.06          536.53           125.09        583.20| |02      12.22           13.32            13.59         13.97| |03      6.72            8.31             7.15          9.04| |04      16.70           31.56            16.61         30.59| |05      8.10            8.05             7.56          7.86| |06      1.32            0.96             1.35          0.95| |07      2.48            3.09             2.55          3.24| |08      10.89           9.82            11.06         9.90| |09      5.44            7.61             5.68          7.67| |10      7.21            11.73             8.24         12.66| As we can see, our IEBins either significantly exceeds the NeWCRFs or achieves on par performance with the latter. [1] ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras, IEEE Transactions on Robotics, 2017 We further evaluate IEBins and NeWCRFs [6] on the single-view reconstruction task. The RMSE (m) metric is used. |Method RMSE (NYU) RMSE (KITTI) || |:---|:---| |IEBins      0.195          1.481| |NeWCRFs 0.205          1.526| As we can see, our IEBins exceeds NeWCRFs by 4.9% and 2.9% on NYU and KITTI datasets, respectively.
Thanks for your reply! Your response answered my queries. From the table you shared, it is evident that the idea works. Although the response to W3 makes sense, it is highly contextual and not an established fact e.g. AdaBins performs best at 256 bins (>> nbins in IEBins). So it is still suggested to rephrase or remove. LocalBins comes very close (equal in 3 out of 6 metrics) and makes the improvements weaker. The visualization can be better in all figures. The results in all cases should be normalized uniformly (e.g. use a dataset global min (e.g. 0.1 meters) and max depth (e.g. 10 meters) for depth maps, and not individual basis). This includes the uncertainty visualization as commented above. In summary, I believe the overall design of the proposed framework is quite nice and extendable. There are missing parts and weaknesses but the design and performance outweigh them. The paper is worth the acceptance.
Decision
Accept (poster)