LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object Detection

Q1: Defend our novelty.

A1: Our research is rooted in a deep understanding of the sparsity characteristics in the LiDAR scenario. We discover two fundamental phenomena (i and ii on Page 5) when it comes to quantization over point cloud data that intrinsically bears sparsity. The choice of the max-min calibration strategy is non-trivial. By selecting maximum/minimum values, we are free from the influence of sparse data distribution. It retains the most effective information and is not affected by the large number of zeros. On the contrary, an entropy-based calibration seeks to preserve data distribution which suffers from abundant zero values and sparse information, rendering much inferior performance (Table 1 and 2). For instance, when the distance range gets beyond 50m, its mAPH suddenly drops from 38.09 to 5.9. Besides, Our proposed max-min doesn't come alone, we apply a simple grid search strategy for better quantization scales. As for TGPL, we pursue a global supervised mechanism to bridge the performance gap caused by the previous per-layer quantization. These three components altogether bring a state-of-the-art performance shown in Table 4. Each component alone may seem simple, but an orchestration of them requires deeper thoughts.

Q2: Claims about point cloud coordinates.

A2: In fact, absolute coordinates are also used in practice. In BEV-based point cloud detectors, such as voxel feature encoding (VFE) and pillar feature encoding (PFE), the Euclidean distance calculated based on the absolute coordinate of the point cloud or the absolute coordinates are also utilized as input features, along with the relative coordinates of voxels or pillars, which is a common operation. Please refer to open-source code https://github.com/tianweiy/CenterPoint/blob/d3a248fa56db2601860d576d5934d00fee9916eb/det3d/models/readers/pillar_encoder.py#L147 or https://github.com/open-mmlab/mmdetection3d/blob/5c0613be29bd2e51771ec5e046d89ba3089887c7/mmdet3d/models/voxel_encoders/voxel_encoder.py#L432.

Q3: Application on other LiDAR detectors.

A3: Thanks for the advice about more LiDAR detectors. We have conducted experiments on FSD, an excellent fully sparse 3D detector, as shown in the table below. Adopting entropy calibration still leads to a significant accuracy drop of -61.50 mAPH/L2. We discover that quantized FSD readily delivers the desired performance while employing a vanilla max-min calibration. Nonetheless, using LiDAR-PTQ can further achieve comparable accuracy to its float counterpart. This demonstrates the effectiveness of LiDAR-PTQ on fully-sparse detector as well. To our best knowledge, the mainstream approaches in industrial applications are still non-fully sparse detectors like VoxelNet[1], PointPillars[2], and CenterPoint[3] because of friendly deployment using off-the-shelf frameworks like TensorRT. We emphasize that our method can be seamlessly integrated to enjoy a significant speed-up, which is vital for the community. In summary, our LiDAR-PTQ is not only limited to specific BEV-based point cloud detectors. Given such generalization ability and practical value for promoting the application of 3D point cloud detectors on edge devices, we believe our method deserves to be shared with our community.

References:

• [1] VoxelNet: End-to-end learning for point cloud based 3D object detection. CVPR 2018.
• [2] Pointpillars: Fast encoders for object detection from point clouds. CVPR 2019.
• [3] CenterPoint: Center-based {3d} object detection and tracking. CVPR 2021.

Q1: Training details

A1: In LiDAR-PTQ, for every quantized layer, the optimized parameter is weight scale $s_w$, weight zero-point $z_w$, activation scale $s_a$, activation zero-point $z_a$ and adaptive rounding value for weight $\theta$. We generally summarize the overall optimization step as follows: in every quantized layer

1. using sparsity-based calibration to get initial $s_w$, $z_w$, $s_a$, $z_a$ and $\theta$.

2. using ${L}_{total}$ to fine-tune the $s_w$, $z_w$, $s_a$, $z_a$ and $\theta$.

The above steps are calculated for every quantized layer in quantized model. More details refer to Algorithm 1 in Sec 3.2.

Q2: Calibrated data percentage.

A2: In WOD dataset, we randomly sample 256 frames point cloud data from the training set as the calibration set. The calibration set proportions is 0.16% (256/158,081) for WOD. Notably, we do not use the ground truth labels for these point cloud data due to we are PTQ method.

Q3: Presumably requires the labelled training data.

A3: We do not require any labeled data to train the quantized model. We use the prediction of FP model to supervise the prediction of quantized model. Employing a method similar to distillation to enhance the quantized model using its float counterpart is effective in PTQ scenarios.

Q4: Gains of TGPL and Adaptive round.

A4: The contribution of TGPL and adaptive rounding is to mitigate the final performance gap and achieve almost the same performance as FP model. There are many advanced 2D PTQ methods (like BRECQ[1], QDROP[2]) that can be used in point cloud detection models quantization. However, they still suffer from performance drops (see Tab 3). In this paper, our motivation is to propose a PTQ method designed for point cloud detection tasks and achieve comparable accuracy with FP model. Although the proposed sparsity-based calibration has shown huge performance gain, there still exists a significant gap compared with the FP model. Therefore, we propose TGPL and adaptive rounding to narrow the final performance gap.

Q5: Open-sourced code.

A5: We are arranging the code and will release it to the community later.

Q6: LEVEL2 in Waymo dataset:

A6: In Waymo Open Dataset (WOD), each ground truth label is categorized into different difficulty levels:

• LEVEL1 (L1): if number of LiDAR points $n$ $\textgreater$ 5 in ground truth boxes.
• LEVEL2 (L2): if number of LiDAR points 1 $\leq n \leq$ 5 in ground truth boxes. When evaluating, L2 metrics are computed by considering both L1 and L2 ground truth. The Waymo official leaderboard ranks based on Mean Average Precision with Heading / LEVEL2 (mAPH/L2) metrics, which is also widely adopted metrics by the community. We have added more details in the manuscript to make it clearer.

Q7: Typos:

A7: We apologize for our error. We have corrected all the typos carefully in the manuscript.

Q8: Page 7, Eq 11:

A8: Yes. In Sec 3.4, the soft-quantized weights is calculated by Eq 10 and Eq 2, and then we calculate ${L}_{total}$ using Eq 11.

References:

• [1] Brecq: Pushing the limit of post-training quantization by block reconstruction. ICLR 2021.
• [2] QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization. ICLR 2022.

Dear Reviewer 3wrT,

Thanks again for the reviews. We expect any further feedback to enhance our paper and address your rest concerns before the deadline. We are desirous to have a productive conversation.

Sincerely,

The Authors

Q1: Add inference performance in Tab.3

A1: Thanks, we have updated it in Tab. 3.

Q2: γ setting in TGPL

A2: In TGPL loss, we set γ as 0.1 to filter predictions from the FP model, and then we select the top K boxes, where we set K as 500. We have updated in manuscript.

Q3: This work like the knowledge distillation from FP model for quantized model?

A3: Yes, employing a method similar to distillation to enhance the quantized model using its float counterpart is effective in PTQ scenarios. To further improve our approach, we have introduced a global distillation scheme called TGPL. Different from the commonly used knowledge distillation paradigm, we did not use any labels of the calibration set in the whole quantization process in LiDAR-PTQ.

Q4: LEVEL_2 (L2) in Waymo dataset

A4: In Waymo Open Dataset (WOD), each ground truth label is categorized into different difficulty levels.:

• LEVEL1 (L1): if number of LiDAR points $n$ $\textgreater$ 5 in ground truth boxes.
• LEVEL2 (L2): if number of LiDAR points 1 $\leq n \leq$ 5 in ground truth boxes.

When evaluating, L2 metrics are computed by considering both L1 and L2 ground truth. The Waymo official leaderboard ranks based on Mean Average Precision with Heading / LEVEL2 (mAPH/L2) metrics, which is also widely adopted metrics by the community.

Q5: Effect for other types of detectors (such as Transform-based).

A5: 1. Thank for your suggestions. Currently, this work primarily focuses on CNN-based Lidar detectors, which remain the mainstream approach in current industrial applications. Besides, we endorse the potential and promising prospects of Transformer-based methods. Combination of Transformer networks and Lidar data will bring further intriguing challenges. We are eager to explore this topic in our future works.

2. Not only limited to the CenterPoint (-Pillar and -Voxel) model, we also evaluated the effectiveness of our method on fully sparse detectors, such as FSD[1], and achieved comparable performance with the FP model. Additionally, we provided the differences between different model architectures, which show LiDAR-PTQ is applicable to differnt cnn-based detectors. | Models | Methods | Bits(W/A) | Mean mAP/L1 | Mean mAPH/L2 | Veh mAP/L2 | Veh mAPH/L2 | Ped mAP/L2 | Ped mAPH/L2 | |-----------|------------|-----------|---------------|----------------|--------------|---------------|--------------|---------------| | FSD[1] | Full Prec. | 32/32 | 73.01 | 70.94 | 70.34 | 69.98 | 73.95 | 69.16 | | FSD[1] | LiDAR-PTQ | 8/8 | 72.84 | 70.73 | 69.95 | 69.62 | 73.85 | 68.95 |

References:

• [1]. Fully Sparse 3D Object Detection, NeurIPS 2022.

Q1: nuScenes dataset experiments

A1: Thank you for your reminder. In the revision, we have updated the performance on the nuScenes [1] dataset. In the nuScenes dataset, we randomly sample 256 frames point cloud data from the training set as the calibration set. The calibration set proportions are 0.91% (256/28,130). Our performance evaluation involves two metrics, average precision (mAP) and nuScenes detection score (NDS). NDS is a weighted average of mAP and other attributes metrics, including translation, scale, orientation, velocity, and other box attributes. As shown in the table below, LiDAR-PTQ sitll shows state-of-the-art performance, consistent with its accuracy on Waymo dataset. Our LiDAR-PTQ achieves almost the same performance as the full precision model on the nuScenes dataset. All experiments have been updated in the supplementary materials.

We hope to resolve your doubts with our best efforts. Please share your thoughts on viewing our reply.

References: [1] Nuscenes: A multimodal dataset for autonomous driving. CVPR, 2020.

Dear Reviewer Dy1B,

Many thanks for your valuable remarks.

To make the best use of the discussion period and to improve our work, we'd like to hear your response to know whether we have addressed your concerns or there are any new questions arise.

We hope to resolve your doubts with our best efforts. Please share your thoughts on viewing our reply.

Regards,

The authors

Q1: Would be similar results on other object detectors?

A1: 1. CenterPoint integrates two milestone works in LiDAR-based BEV detection, VoxelNet and PointPillars as CP-Pillar and CP-Voxel. In particular, CP-Pillar and CP-Voxel have different network designs. The CP-Pillar model is a fully dense convolutional network, while the CP-Voxel model includes SPConv and dense convolution. Our results on CenterPoint (-pillar and -voxel) demonstrate that:

• Lidar-PTQ is applicable to pillar-based and voxel-based detectors.
• Lidar-PTQ is applicable to SPConv and dense convolution operations.

2. Except CP-Pillar and CP-Voxel models, to evaluate the generalization ability of our method, we also evaluate our LiDAR-PTQ on FSD, a fully sparse 3D detector. The experiments demonstrate that our method is also applicable to fully sparse detectors.

Q2: Would be similar results on other object applications, e.g., segmentation?

A2: Thanks for your constructive suggestions.

1. For a timely response, we have not enough time to provide experiments for point cloud segmentation task due to the original intention of this paper is to focus on the LiDAR-based BEV detection task. LiDAR point cloud segmentation is another topic, which beyond the original scope of this study.
2. Furthermore, we are actively conducting experiments in point cloud segmentation task, and if time permits, we will provide the related results before the rebuttal deadline to better evaluate our LiDAR-PTQ method.

Q1: Would be similar results on other applications using Lidar data, e.g., segmentation.

A1: Given your prior constructive suggestions, we conducted experiments on SemanticKITTI [1] dataset for point cloud segmentation to further evaluate the generalization of LiDAR-PTQ. Specifically, we utilize SPVNAS [2] as our baseline, which is a representative work in point cloud segmentation. As shown in table below, adopting entropy calibration leads to a significant accuracy drop of 18.09 mIOU. As for a vanilla max-min calibration, there is still a performance drop 2.64 mIOU for quantized SPVNAS. However, LiDAR-PTQ can further achieve comparable accuracy to its float counterpart. This demonstrates the effectiveness of LiDAR-PTQ on point cloud segmentation tasks as well. For more detailed performance of different classes refer to Tab. 2.

We hope to resolve your doubts with our best efforts. Please share your thoughts on viewing our reply.

References:

1. SemanticKITTI: A dataset for semantic scene understanding of lidar sequence, ICCV 2019.
2. Searching efficient 3d architectures with sparse point-voxel convolution, ECCV 2020.

Dear reviewers,

Could you please read the authors' responses and give your feed back? The period of Author-Reviewer Discussions is Fri, Nov 10 – Wed, Nov 22.

Many thanks,

AC