Summary
One main challenge of monocular 3D object detection models is the lack of depth information from RGB images. The authors proposed a metric learning scheme to encourage the model to extract depth-discriminative features. Based on the presented theoretical results, the authors proposed a quasi-isometric loss and an object-wise depth map loss to supervise the model. Quantitative results on benchmark datasets supported the main arguments of the paper and sufficient ablation study experiments were conducted.
Strengths
1. The motivation and problem setting are explained clearly. References were provided to demonstrate why depth is key to improve current monocular 3D methods and how depth-discriminative features would help.
2. The proposed quasi-isometric loss seems novel and effective. To prevent the depth-discriminative loss from damaging the non-linearity of the natural manifold, the authors adopted a distance-preserving condition. Despite extra hyper-parameters introduced, the proposed method is effective in general under various settings.
3. Quantitative results on two benchmark datasets using multiple baseline models demonstrated the effectiveness of the method. Results showed that the proposed depth-discriminative is an effective approach to assist monocular 3D methods.
Weaknesses
1. The high-level idea of the proposed approach resembles previous contrastive learning approaches [1]. I could imagine adding depth-based feature contrastive losses to the baseline loss. Would that work? What would be the advantage of the proposed approach compared to contrastive losses?
2. The authors claimed that directly learning quasi-isometry would hurt sub-tasks. Despite such trade-offs being common in deep learning, I think this problem is a bit understudied. For instance, object features from objects that differ a lot in distance should be easily discriminated against. Would larger $K$ help with a larger $\epsilon$? Or is this design related to hard negative sampling?
3. Figure 2 seems interesting. However, I assume the pairs are also defined by the given $K, B, \epsilon$, which makes it less convincing.
References:
1. D. Neven et al. Towards End-to-End Lane Detection: an Instance Segmentation Approach.
Questions
1. What is “Ours” in supplementary Table 3? Only $\mathcal{L}\_{qi}$ or both $\mathcal{L}\_{qi}$ and $\mathcal{L}\_{obj}$? I assume when ablating $K, B, \epsilon$, you should not involve $\mathcal{L}\_{obj}$ in the comparison?
2. Since there is an auxiliary head for depth estimation, is the estimated depth more accurate with the quasi-isometric loss? It would be good to compare the trade-off between depth estimation and “other sub-tasks” when $\epsilon$ changes.
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.
Limitations
1. The work can be improved by comparing with methods with similar ideas (see weakness 1).
2. It is a bit unclear how different feature spaces look like under different hyper-parameter settings. It would be good if more analysis are presented (see weakness 2).