Going Deeper with General and Specific Inductive Bias for Real-Time Stereo Matching

1. The writing needs improving. Introducing the IB concept makes it hard to understand what the authors claim. I wonder why it needs to introduce IB instead of simply exploring scale invariance and long-range dependencies.
2. The results of the KITTI benchmarks are not impressive. Among real-time methods, we find CGI-Stereo, HITNet, and Fast-ACVNet+ get better performances.
3. There is a lack of evaluating the generalization ability in common settings. This paper claims UStereo has robust generalization, however, the common setting for generalization evaluation is on the KITTI, Middlebury, and ETH3D datasets.

• The claim of the proposed method is robust in generalization is not well justified. It finetunes the pre-trained model in the target dataset and is evaluated in the target dataset. In [A], they claim to be robust in generalization as it is not finetuned in the target dataset.

• The ablation study is not well-written. It is difficult to follow the decision process of each module (DSF & MLDC). It is better to focus on each module separately. It is recommended to follow Table 2 in [A] to show better ablation study. It is unclear why the softmax normalization is replaced by Sigmoid function to validate the long range dependencies. Wouldn't it better to remove the MDLC module for the ablation study as in Exp3?

• The goal of the specific IB is to perform the deep supervision in the lower resolution so that the distribution of the lower-level cost volume can be supervised by the distribution of ground truth disparity. The idea of deep supervision is not new because it has been done in [B] as well. However, it is unclear how the proposed additional supervised loss can lead to better performance.

• The performance improvement seems to be incremental, while the justification of the proposed method is lacking due to unclear experimental results. More exhaustive ablation study is required to justify the performance, especially comparison with various layer fusion methods, global and local attention stereo matching methods.

• The motivation of the problem is unclear. There is no experimental based justification that the current stereo matching method performance is insufficient due to lack inductive bias in the network design. Selecting the state-of-the-art methods and showing the problem quantitatively might be better approach to show the motivation.

Additional references

• [A] Domain Generalized Stereo Matching via Hierarchical Visual Transformation, CVPR 2023
• [B] CFNet: Cascade and Fused Cost Volume for Robust Stereo Matching, CVPR 2021

Very poorly written paper, should be rejected for now based on writing quality alone. Could have a contribution but would need to be rewritten. Suggestions.

1. Review the literature on deep learning for stereo matching
2. Explain how your chosen inductive bias has not been seen in the literature
3. Note the use of attention mechanisms to model long range dependencies
4. Introduce your model of stereo. Explain the role of scale etc.
5. Showcase your results.
6. Drop the real-time aspect.

The paper is very hard to read, both because of the writing quality and the lack of clarity. Some sentences make very little sense, possibly because this. As an exemple (p.6):

The learning process would make internal relationships vague unlike Eq.9

Figure 1, 2 and 3 are not clear enough to help understand the architecture.

The ablation study does not clearly illustrate how the proposed modules actualy get better performance. Also, many decisions in the architecture are not justified and would deserve to be part of the ablation study (ex: using Top-5 for disparity regression)

Eq.12 suggests that the Max of a Softmax, which is a probability, will be minimized. Can this be justified?

Results : Table 1 provide performance on KITTI12 and KITTI15, but the paper results are not available in the online dashboard. This is very important and must be fixed if the results are to be taken seriously.

Also, Table 1 result for BGNet+ KITTI2015 D1 result is not 2.91, but 2.19. Moreover, many results of Table are cited from published papers, but the actual online results are better (see CoEx and FastACV). Even if this is not a major problem, this disrepency should be mentioned by the authors.

Table 4 is hard to understand, as well as the coding (32L,8L,32F, etc.) used for deep extra supervision.

The two extra statistics indicators ($\delta$ 50%top and $\delta$ 50%mean) are not well justified. Maybe an ablation study would help justify them. The pseudocode (p.17) is not useful, as these indicators are very easy to compute.

Many Spelling and formulations problems (here are a few, but there are many more):

• p.1 invarience
• p.1 for almost vision tasks
• p.2 knowledgement
• p.3 wildly -> widely
• p.5 our contemplation
• ...

There are major issues with the paper writing. While the proposed ideas may sound, it used a rather indirect way to make the presentation.

The evaluations are insufficient. It did not perform sufficient evaluation in demonstrating the superority of the proposed solution in terms of accuracy, running time and memory.