Depth-discriminative Metric Learning for Monocular 3D Object Detection

Monocular 3D object detection poses a significant challenge due to the lack of depth information in RGB images. Many existing methods strive to enhance the object depth estimation performance by allocating additional parameters for object depth estimation, utilizing extra modules or data. In contrast, we introduce a novel metric learning scheme that encourages the model to extract depth-discriminative features regardless of the visual attributes without increasing inference time and model size. Our method employs the distance-preserving function to organize the feature space manifold in relation to ground-truth object depth. The proposed (K, B, eps)-quasi-isometric loss leverages predetermined pairwise distance restriction as guidance for adjusting the distance among object descriptors without disrupting the non-linearity of the natural feature manifold. Moreover, we introduce an auxiliary head for object-wise depth estimation, which enhances depth quality while maintaining the inference time. The broad applicability of our method is demonstrated through experiments that show improvements in overall performance when integrated into various baselines. The results show that our method consistently improves the performance of various baselines by 23.51% and 5.78% on average across KITTI and Waymo, respectively.

Paper

Similar papers

Peer review

Reviewer eeNQ8/10 · confidence 5/52023-07-01

Summary

This work focuses on the monocular 3D object detection task. As many works indicated, depth estimation is the bottleneck of this task, and the authors propose the apply metric learning to improve the accuracy of the depth estimation sub-task. The proposed metric-learning-based loss encourages the model to extract depth-discriminative features regardless of the visual attributes without increasing inference time and model size. Extensive experiments on the KITTI3D and Waymo Open datasets demonstrate the effectiveness of the proposed loss fuction.

Strengths

1. I like the idea that applying metric learning to improve the accuracy of depth estimation. Theoretically, this method has a good generalization ability and is easy to be embedded in other baseline models. 2. Extensive experiments and good performance. The authors conduct lots of experiments on the KITTI3D and the large-scale Waymo Open datasets. Besides, they test their metric learning-based loss function on several baseline models, and the improvements across several settings demonstrate the effectiveness of their proposed model. 3. The main idea is easy to follow and this paper is well-organized. This paper has many mathematical details that scholars in this field may not be familiar with, but it still presents them clearly and systematically.

Weaknesses

1. In lines 68-69, the authors claim their proposed work is 'the first approach that applies metric learning to monocular 3D object detection.' In fact, there is another work [1] that discussed how to apply metric learning in monocular 3D object detection with a focus on dimension estimation. This claim should be modified or removed. 2. Although this work provides lots of experiments to show the effectiveness of their method, lots of them are based on the same foundation model, i.e. CenterNet (and one of them is based on the transformer-based pipeline). However, there are still some other popular detection pipelines such as the BEV paradigm. It will be better to validate the effectiveness of more pipelines. 3. This work conducts experiments on the large-scale Waymo dataset, which is great. However, I still want to see the performance on the nuScenes dataset, because the nuScenes benchmark is dominated by another detection pipeline and adopts different evaluation metrics. Evaluating on this dataset can further show the effectiveness and generalization ability of the proposed method. Overall, this paper proposes a novel and effective metric-learning-based loss function. I tend to accept this work and I can further improve my rating if the authors can show the generalization ability of this work in more settings. [1] Dimension Embeddings for Monocular 3D Object Detection, Zhang et al., CVPR'22

Questions

See weaknesses.

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

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

4 excellent

Contribution

3 good

Limitations

See weaknesses.

Reviewer tr2v6/10 · confidence 4/52023-07-06

Summary

This paper introduces a novel metric learning scheme for extracting more depth-discriminative features in the monocular 3D object detection task. A distance-preserving function is adopted to build the relation between feature space and the ground-truth object depth. The authors propose a quasi-isometric loss to adjust the distances among object descriptors. Furthermore, an auxiliary head for object-wise depth estimation is used during training, enhancing depth quality while maintaining the inference time. The experiments show that the proposed method can improve the performance significantly of various baselines on KITTI and Waymo datasets.

Strengths

- This paper introduces a metric learning method to improve the discrimination of object descriptors according to their depth information. The scheme is simple yet effective, which maintains the geodesic distance of depth information in feature space. The quasi-isometry under the local distance-preserving condition mitigates the negative damage to the non-linearity of the natural manifold. It minimizes the influence on other tasks. - The quasi-isometric loss is designed to arrange condition-violated samples. The samples far from the quasi-isometric distance are pulled closer. The samples too close in feature space are pushed away. The loss focuses on hard samples which avoids many unnecessary computations while ensuring the isometric objective. - The proposed method is plug and play. It can easily integrate into various monocular 3D object detection baselines and brings significant improvement.

Weaknesses

- I am curious about the efficiency of the proposed method during training. The calculations of relative distances between objects in both original and feature space would be enormous. Besides, the complexity of calculating the distance matrix grows with the square of the number of objects. Therefore, the proposed method may be not suitable for the scenario with a large number of objects during training. The paper does not discuss this problem. - The object-wise depth prediction head is trained for estimating more accurate depth, avoiding significant errors that existed in the center depth prediction head. Why not directly use the prediction of object-wise depth head for a more accurate depth estimation during inference?

Questions

See weaknesses.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors do not address the limitations of their work. This method may be limited by the number of objects.

Reviewer nNTt7/10 · confidence 4/52023-07-07

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.

Soundness

3 good

Presentation

3 good

Contribution

3 good

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).

Reviewer 6BaQ7/10 · confidence 4/52023-07-08

Summary

The paper proposed a new approach to monocular 3D object detection. The critical contribution of the work is the application of metric learning to improve depth estimation and an additional head with auxiliary depth prediction. The resulting approach improves 3D object detection accuracy without increasing inference time, model size, or additional data. The proposed method utilizes a metric learning scheme that preserves the geodesic distance between depth information and the feature space. This scheme encourages the model to extract depth-discriminative features without negatively impacting other non-depth tasks (object size, type and etc.). An auxiliary head is also introduced to enhance depth estimation, adapting from [20]. This auxiliary head improves the quality of depth estimation without impacting inference time, ensuring efficient performance. The experimental evaluations conducted on the KITTI and Waymo datasets demonstrate the effectiveness of the proposed method. The results consistently show solid improvements in performance across various monocular 3D object detection methods. ------ Updated my score after rebuttal.

Strengths

- The paper is well-written and easy to follow. The authors did a great job describing the quasi-isometric concept, its mathematics, and how the proposed loss enforces them. - The paper is well-intuitive. The different asks in 3D object detection can create conflict in feature extraction, negatively impacting each task. The idea of using metric learning to provide a feature manyfold that balances the discrimination of depth and local structure for other tasks is promising. - The related work section is comprehensive, covering different related areas. - The experiment demonstrates the strength of the proposed method and generalizability on different baselines and datasets. Ablation analysis is done on the two key components to show their effectiveness and contribution.

Weaknesses

- The paper provides two essential components contributing to the final improvement. But I don't think the second auxiliary head contains a lot of novelty compared to related works such as MonoCon. This limits the novelty of the paper to some extent. - It's great that the paper compared to SupCR in ablation analysis. However, there is a missing discussion from a more theoretical standpoint on why the proposed loss is better than contrastive loss. In the end, both of the losses seem to be able to maintain a certain local distance for the other tasks. - Part of the parameter-sensitive analysis should be moved to the main paper.

Questions

- Can I ask why not use nuScenes, which seems to be more popular in monocular 3D detection in the past few years? - How does P correspond to the feature map h in section 3.2? Is P the feature space that the map predicts? Or is a different one going through some MLP? - Is there any systematic way for parameter choosing? Even just a starting set? - Have the authors tried other metric learning methods that didn't seem to work?

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.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The parameter of the loss seems to be quite sensitive to the final performance. Would it be helpful to provide guidelines or an algorithm for parameter initialization?

Reviewer Wa1k6/10 · confidence 3/52023-07-13

Summary

This paper proposes a metric learning scheme to learn depth-discriminative features for object depth prediction, which helps improve the overall task of monocular 3D object detection, without negatively impacting the performance of the other sub-tasks (e.g., object class, bounding box size) wherein. Specifically, they employ a distance-preserving function and the proposed ($K, B, ε$)-quasi-isometric loss to arrange the feature space manifold in accordance with ground-truth object depth, while preserving the non-linearity of the natural feature manifold. They also introduce an auxiliary head (in training) for object-wise depth estimation to enhance the depth quality. Experiments on datasets KITTI and Waymo show that the proposed method can be incorporated into several 3D object detection backbones for improvement.

Strengths

- The proposed $(K, B, ε)$-quasi-isometric loss is mathematically explained to help learn depth-discriminative features. - Experiments on KITTI and Waymo show the effectiveness of the proposed method, and the ablation studies are well-designed for verification including several backbones and comparison with other SOTA (e.g., SupCR). - The paper is well written, including the problem definition, the purpose, and the background. The equations and the detailed notations help to understand the method.

Weaknesses

See **Questions** below.

Questions

- Are those hyper-parameters $K$, $B$ and $\epsilon$ easy to find in pratice? Will different backbone architectures require different setups of $K$, $B$, and $\epsilon$? - According to Eq 6, the proposed Quasi-isometric loss is implemented as the contrastive loss. 1) Are there any restrictions on the absolute number and the relative ratio of positive/negative samples/anchors? 2) Is there any direct comparison between the proposed Quasi-isometric loss with the naive contrastive loss for metric learning in the task of monocular 3D object detection? - It would be helpful to show some visualization results of the learned features.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

- N/A for the limitations. - I suggest the authors provide the failure cases (and the corresponding explanations) of the proposed losses on monocular 3D object detection.

Reviewer eeNQ2023-08-11

Final Rating

I appreciate the additional experiments which further confirm the effectiveness and generalization ability of the proposed method. Based on the solid theoretical analysis and good results, I'd like to change my score to 'strong accept'.

Reviewer nNTt2023-08-13

Final Rating

Thank the authors for preparing the additional experimental results. They addressed my concerns and I believe the results and discussions would a good addition to the paper. I think this is a strong submission and recommend "7: Accept".

Reviewer tr2v2023-08-14

The reply addresses my concerns

Thanks for your reply. The feedback has addressed my concerns. After reading other reviews and the rebuttal materials, I lean to acccept this paper.

Reviewer 6BaQ2023-08-21

I appreciate the authors' effort in the rebuttal—the rebuttal address most of my concerns. I want to increase my rating based on the explanation and the new results.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC