the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
A Lightweight BEV Perception Optimization Framework for Real-Time 3D Object Detection in Occluded Scenes
Abstract. Real-time 3D object detection in occluded scenes is a core challenge for pure-vision autonomous driving perception. Existing dense Bird's Eye View (BEV) detection methods suffer from redundant backbone parameters, inflexible fixed sampling strategies, and coarse-grained temporal modeling, making it difficult to simultaneously satisfy detection accuracy and real-time latency requirements on embedded vehicle platforms. This paper proposes a lightweight BEV perception optimization framework for occluded scenes. An occlusion-aware adaptive voxel feature sampling module dynamically switches between fast ray projection and deformable attention paths according to scene complexity. A temporal grouping fusion module based on Res2Net principles performs grouped cross-frame fusion of consecutive BEV feature maps without introducing additional learnable parameters. A two-stage LiDAR-to-camera knowledge distillation scheme with a geometric compensation module transfers depth geometry knowledge during training while incurring zero inference overhead. Experiments on the nuScenes dataset demonstrate that the proposed method achieves 38.7% mAP and 51.3% NDS under the lightweight configuration, with an inference latency of 38.2 ms and 26.2 FPS on NVIDIA Tesla T4, ranking highest in NDS among comparable methods.
- Preprint
(5167 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on ms-2026-168', Anonymous Referee #1, 27 Sep 2026
-
AC1: 'Reply on RC1', Yi Zhang, 22 Nov 2026
We thank the reviewer for the positive assessment of the proposed lightweight BEV framework and its accuracy–efficiency trade-off. We reply the comments below.
Comment 1: The manuscript is motivated primarily by the occlusion problem. The authors are encouraged to provide more rigorous experiments to verify whether the proposed complexity metric effectively reflects the degree of occlusion and to quantify the performance improvements under different occlusion levels.
Response:
Thank you for your comments! We agree that scene complexity and occlusion severity are not equivalent. In Section 3.3.1, the score s = αμ + (1 − α)σ is computed from the mean and standard deviation of normalized BEV features; it is not an annotation-based measure of occlusion. The existing ablation results show that adding adaptive sampling increases NDS from 42.6% to 46.9%. Compared with always-on deformable sampling, the adaptive scheme reduces latency by 18.2 ms at a cost of 0.7 NDS percentage points. These results support an accuracy–latency trade-off but do not establish a direct correlation between the score and occlusion severity or quantify performance across occlusion levels. We will therefore describe the score as a feature-based proxy for scene complexity and qualify the associated occlusion-related claims.Comment 2: The claim regarding real-time embedded deployment appears stronger than what is currently supported by the experimental evidence, as inference efficiency is evaluated only on an NVIDIA Tesla T4 rather than on an embedded/mobile platform. The power consumption of the T4 is also substantially higher than that of mobile computing platforms with a typical power budget of approximately 20 W. Evaluation on representative embedded hardware would therefore strengthen the deployment-related claims.
Response:
Thank you for your comments! The reported latency of 38.2 ms (26.2 FPS) was measured on a single NVIDIA Tesla T4 with batch size 1 and an input resolution of 256 × 704 (Section 4.5). These results characterize inference performance on the tested GPU and do not establish speed or power efficiency on a low-power embedded platform. As embedded-device measurements are not available, we will restrict the claim to real-time inference on the T4 and present embedded deployment as a potential application requiring separate validation.Comment 3: The manuscript is somewhat verbose in places. For example, Section 5 repeats much of the numerical evidence already presented in Sections 4.3–4.5. In addition, the Introduction could be streamlined by removing detailed descriptions of prior studies that are subsequently repeated in Section 2.
Response:
Thank you for your comments! The Introduction repeats some literature descriptions from Section 2, while Section 5 restates numerical results already presented in Sections 4.3–4.5. We will shorten the Introduction to emphasize the research motivation, gap, and contributions, retain detailed literature comparisons in Section 2, and focus the Discussion on the interpretation and limitations of the findings rather than repeating the results.Comment 4: The readability and visual presentation of Figures 2, 4, and 5 could be improved.
Response:
Thank you for your comments! We will improve the resolution and legibility of the indicated figures, including text size, line clarity, legends, and the presentation of qualitative comparisons, while keeping figure labels and captions consistent.Citation: https://doi.org/10.5194/ms-2026-168-AC1
-
AC1: 'Reply on RC1', Yi Zhang, 22 Nov 2026
-
RC2: 'Comment on ms-2026-168', Anonymous Referee #2, 11 Oct 2026
The manuscript proposes a lightweight BEV perception framework for real-time 3D object detection in occluded scenes by combining adaptive sampling, temporal fusion, and LiDAR-to-camera knowledge distillation. The topic is relevant and the experimental results are promising. However, several issues should be addressed before publication.
Major Comments
- The occlusion-aware mechanism needs stronger validation. The scene complexity score is defined using the mean and standard deviation of BEV features, but the manuscript does not demonstrate that this score is directly related to occlusion severity. More quantitative evidence is needed to show that the adaptive sampling strategy is truly occlusion-aware.
- The adaptive sampling pipeline is not sufficiently clear. The manuscript states that the scene complexity score is calculated from the current BEV feature map, while this score is also used to determine how the BEV feature map is generated. The authors should clarify the exact processing order and avoid possible circular dependency in the framework description.
- Some claims are inconsistent with the reported results. The manuscript states that the temporal grouping fusion introduces no additional learnable parameters, but the method contains several convolutional layers and Table 3 shows an increase in model parameters. This statement should be corrected.
- The fairness of the baseline comparison should be clarified. The authors should state whether the results of the comparison methods were reproduced under the same settings or directly taken from the literature. In addition, EdgeNeXt-S should not be described as a “ResNet-50 backbone configuration.”
- The latency analysis requires further explanation. The manuscript reports that knowledge distillation reduces inference latency from 43.5 ms to 38.2 ms, although the teacher model is removed during inference. The authors should explain the mechanism responsible for this latency reduction and provide more rigorous timing results.
- The ablation study only demonstrates incremental rather than independent contributions. The modules are added sequentially in Table 3. Therefore, the current results mainly show the incremental effect of each module under a fixed addition order. The wording should be revised, or additional leave-one-out experiments should be provided.
Minor Comments
- “Reasoning efficiency” should be replaced by “inference efficiency.”
- The equation numbering should be carefully checked, as several equation numbers appear duplicated. ms-2026-168
- The qualitative discussion should distinguish between object occlusion and adverse weather or nighttime conditions.
- The resolution and readability of Figs. 1, 4, and 5 should be improved.
- The manuscript would benefit from further English-language polishing.
Citation: https://doi.org/10.5194/ms-2026-168-RC2 -
AC2: 'Reply on RC2', Yi Zhang, 22 Nov 2026
We thank the reviewer for recognizing the relevance of the topic and the potential of the reported results. We appreciate the specific concerns and respond to them below.
Comment 1: The occlusion-aware mechanism needs stronger validation. The scene complexity score is defined using the mean and standard deviation of BEV features, but the manuscript does not demonstrate that this score is directly related to occlusion severity. More quantitative evidence is needed to show that the adaptive sampling strategy is truly occlusion-aware.
Response:
Thank you for your comments! We agree that the current evidence does not directly establish an association between the complexity score and occlusion severity. As defined in Section 3.3.1, s = αμ + (1 − α)σ is a feature-based scene-complexity score, not a measured occlusion ratio. The existing results show an increase in NDS from 42.6% to 46.9% when adaptive sampling is added and an 18.2 ms latency reduction relative to always-on deformable sampling, with 0.7 percentage points lower NDS. These findings support adaptive computation but not validated occlusion-level recognition. We will make this distinction explicit and avoid presenting the score as a direct measure of occlusion severity.Comment 2: The adaptive sampling pipeline is not sufficiently clear. The manuscript states that the scene complexity score is calculated from the current BEV feature map, while this score is also used to determine how the BEV feature map is generated. The authors should clarify the exact processing order and avoid possible circular dependency in the framework description.
Response:
Thank you for your comments! We agree that the current description is ambiguous. Section 3.1 describes complexity scoring as occurring alongside BEV generation, Section 3.3.1 takes the current BEV map as the scoring input, and Section 3.3.2 uses that score to select the BEV sampling path. As written, these steps do not establish a non-circular execution order. We will verify the actual implementation and explicitly identify which features are available before the sampling decision and how the selected path is then executed. We will not imply that the output of a selected path is required to select that same path.Comment 3: Some claims are inconsistent with the reported results. The manuscript states that the temporal grouping fusion introduces no additional learnable parameters, but the method contains several convolutional layers and Table 3 shows an increase in model parameters. This statement should be corrected.
Response:
Thank you for your comments! The temporal grouping fusion module contains trainable convolutional layers, and weight sharing does not make their parameter cost zero. In Table 4-3, adding this module increases the parameter count from 23.1M to 24.6M and latency from 40.2 to 43.5 ms, while NDS improves from 46.9% to 49.7%. We will replace the incorrect claim of "no additional learnable parameters" with a statement that grouping and weight sharing limit, rather than eliminate, the additional parameter and computational costs.Comment 4: The fairness of the baseline comparison should be clarified. The authors should state whether the results of the comparison methods were reproduced under the same settings or directly taken from the literature. In addition, EdgeNeXt-S should not be described as a "ResNet-50 backbone configuration."
Response:
Thank you for your comments! Section 4.2 identifies EdgeNeXt-S as the backbone of the lightweight model, so calling it a "ResNet-50 backbone configuration" in Section 4.3 is incorrect. The current comparison tables also do not distinguish clearly between reproduced results and values taken from published sources. We will clarify the provenance and relevant settings of each baseline after checking the reported sources, and describe comparisons with R50 models as cross-backbone reference comparisons rather than matched-backbone experiments. We will also avoid unsupported ranking claims: for example, Table 4-1 reports 51.3% NDS for our EdgeNeXt-S model and 54.0% for SOLOFusion-R50.Comment 5: The latency analysis requires further explanation. The manuscript reports that knowledge distillation reduces inference latency from 43.5 ms to 38.2 ms, although the teacher model is removed during inference. The authors should explain the mechanism responsible for this latency reduction and provide more rigorous timing results.
Response:
Thank you for your comments! We agree that the current explanation is not adequately supported. The teacher model and training-only distillation losses are not used during student inference (Sections 3.5–3.6), so their removal cannot by itself explain a reduction in the student's forward-pass latency. Table 4-3 reports 43.5 ms and 38.2 ms for configurations (e) and (f), but the current manuscript provides no runtime breakdown establishing the cause of this difference. In particular, the claim that distillation reduces the number of depth-estimation iterations is not substantiated by the method description. We will remove this causal explanation and report the latency values without attributing the difference to distillation itself.Comment 6: The ablation study only demonstrates incremental rather than independent contributions. The modules are added sequentially in Table 3. Therefore, the current results mainly show the incremental effect of each module under a fixed addition order. The wording should be revised, or additional leave-one-out experiments should be provided.
Response:
Thank you for your comments! Table 4-3 reports cumulative configurations (a)–(f), with modules added in a fixed sequence. Thus, the differences between adjacent configurations quantify incremental effects under that sequence, not independent contributions free from module interactions. We will change the description in Section 4.4 accordingly and avoid claiming that the cumulative ablation establishes the independent contribution of every module.Minor Comment 1: "Reasoning efficiency" should be replaced by "inference efficiency."
Response:
Thank you for your comments! We will use "inference efficiency" in the Section 4.5 heading, the Table 4-4 title, and the relevant text.Minor Comment 2: The equation numbering should be carefully checked, as several equation numbers appear duplicated.
Response:
Thank you for your comments! We will check and correct the equation numbering throughout the manuscript and ensure that all in-text references are consistent.Minor Comment 3: The qualitative discussion should distinguish between object occlusion and adverse weather or nighttime conditions.
Response:
Thank you for your comments! Section 4.6 presents qualitative examples under different weather and lighting conditions, which should not be treated as direct evidence of object occlusion. We will distinguish geometric occlusion caused by other objects from visibility degradation caused by rain, low light, or nighttime conditions, and qualify the corresponding interpretations.Minor Comment 4: The resolution and readability of Figures 1, 4, and 5 should be improved.
Response:
Thank you for your comments! We will improve the image quality, labeling, and readability of these figures and ensure consistency between the figures and their captions.Minor Comment 5: The manuscript would benefit from further English-language polishing.
Response:
Thank you for your comments! We agree and will polish the English throughout the manuscript, with particular attention to grammar, terminology, clarity, and consistency.Citation: https://doi.org/10.5194/ms-2026-168-AC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 391 | 105 | 42 | 538 | 61 | 82 |
- HTML: 391
- PDF: 105
- XML: 42
- Total: 538
- BibTeX: 61
- EndNote: 82
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript presents a practical lightweight BEV framework that achieves a reasonable trade-off between accuracy and efficiency. The scene-complexity-driven adaptive switching between inexpensive ray projection and deformable sampling is interesting and has potential practical value for engineering applications.
Some minor suggestions for revision:
1. The manuscript is motivated primarily by the occlusion problem. The authors are encouraged to provide more rigorous experiments to verify whether the proposed complexity metric effectively reflects the degree of occlusion and to quantify the performance improvements under different occlusion levels.
2. The claim regarding real-time embedded deployment appears stronger than what is currently supported by the experimental evidence, as inference efficiency is evaluated only on an NVIDIA Tesla T4 rather than on an embedded/mobile platform. The power consumption of the T4 is also substantially higher than that of mobile computing platforms with a typical power budget of approximately 20 W. Evaluation on representative embedded hardware would therefore strengthen the deployment-related claims.
3. The manuscript is somewhat verbose in places. For example, Section 5 repeats much of the numerical evidence already presented in Sections 4.3–4.5. In addition, the Introduction could be streamlined by removing detailed descriptions of prior studies that are subsequently repeated in Section 2.
4. The readability and visual presentation of the Figure 2, 4, and 5 could be improved.