BEVDet: High-performance Multi-camera 3D Object Detection in Bird-Eye-View
1. Paper overview
In one sentence: BEVDet makes camera-to-BEV detection effective through separate augmentation in each view space, then improves small-object suppression and pooling efficiency. architecturenetworkprojectionaugmentation-ablationscale-nmspooling
| At a glance | What to know |
|---|---|
| Research problem | Source description The paper asks whether multi-camera 3D detection can share the BEV representation used by semantic mapping while improving accuracy and latency. Geometry-related targets are naturally expressed on the ground plane, but simply assembling existing modules overfits. The challenge is therefore both representation design and regularizing the modules after projection. architectureaugmentationaugmentation-ablation |
| Core mechanism | Source description A modular camera detector combines an image encoder, Lift-Splat-Shoot view transformation, a BEV encoder, and the first-stage CenterPoint head, without relying on LiDAR depth supervision. architecturenetwork |
| A key reported result | Multi-camera 3D detection on nuScenes validation: Tiny: 31.2% mAP, 39.2% NDS, 215.3 GFLOPs, 15.6 FPS at 704×256. Base: 39.3%, 47.2%, 2,962.6 GFLOPs, 1.9 FPS at 1600×640. mAP and NDS (higher better); GFLOPs and FPS. Official 700/150/150 scene split; six cameras, ten classes. Table 2 unmarked camera rows, without test-time augmentation or ensemble. FCOS3D: 29.5% mAP, 37.2% NDS, 2,008.2 GFLOPs, 1.7 FPS at 1600×900. PGD: 33.5% mAP, 40.9% NDS, 1.4 FPS. Tiny improves FCOS3D by 1.7 mAP and 2.0 NDS percentage points. Comparisons use different resolutions and backbones; monocular FPS is divided by six, with acknowledged batching headroom. They do not isolate architecture alone. protocolvalidationtiming |
| Reading caution | Reader analysis Attribute prediction remains weaker: Base's validation mAAE is 0.191 versus FCOS3D's 0.170, where lower is better. The proposed explanation—that appearance cues are easier to retain in image view—is a conjecture. Joint mapping/detection and driving execution are not evaluated; multi-task learning remains future work. validationattributes-future |
Core contributions
- Source description
A modular camera detector combines an image encoder, Lift-Splat-Shoot view transformation, a BEV encoder, and the first-stage CenterPoint head, without relying on LiDAR depth supervision. architecturenetwork
- Source description
Separate geometric augmentation in image and BEV spaces, category-dependent Scale-NMS, and auxiliary-index pooling develop accuracy and inference efficiency beyond the initial module assembly. augmentationscale-nmspooling
Figure 1. Camera features become a common BEV representation before object detection. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the camera stack at the lower left and follow the arrows into the image encoder. The C and U markers mean concatenation and upsampling. In the view transformer, the upper branch predicts depth-classification scores, while the lower branch supplies image features; O combines them by an outer product. The camera model locates the resulting features in 3D, and P pools them into the BEV map. Follow the pink feature hierarchy through the BEV encoder to the detection head. The illustrated dimensions describe Tiny; Table 1 on page 6 defines the other configurations. architecturenetworkaugmentation
What it supports. The depth branch connects image appearance to a geometrically organized feature map. The head then predicts object properties from BEV features. This modular construction explains why BEV encoding and the regularization applied after projection can be studied separately from image encoding.
Where the evidence stops. The depicted point cloud is generated by feature lifting, not a LiDAR input. Section 3.1 specifies the first-stage CenterPoint head only. The diagram contains neither an action-conditioned rollout nor an executed driving policy.
2. Motivation
2.1 The problem and the proposed response
The paper asks whether multi-camera 3D detection can share the BEV representation used by semantic mapping while improving accuracy and latency. Geometry-related targets are naturally expressed on the ground plane, but simply assembling existing modules overfits. The challenge is therefore both representation design and regularizing the modules after projection. architectureaugmentationaugmentation-ablation
2.2 What this reading follows
BEVDet asks a practical question: if object position and motion are described on the ground plane, why perform all visual reasoning in camera coordinates? Its answer lifts image features into a shared bird's-eye-view map and applies a conventional 3D detection head. The interesting part is what makes this assembly work. Geometrically correct projection cancels image transformations, so the BEV network needs its own augmentation. Small boxes also need a different suppression strategy, and pooling needs an implementation suited to parallel computation. This reading follows those three choices through the original diagrams and nuScenes ablations, while keeping scene perception separate from demonstrated autonomous control. architecturenetworkprojectionaugmentation-ablationscale-nmspooling
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is Components of WAMs. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational representation/state-estimation classification is supported by a camera-to-BEV perception pipeline. Its outputs are current object estimates, including velocity, rather than future states conditioned on actions. No policy, joint future/action predictor or inverse-dynamics action extractor is present. Consequently the world-action architecture, prediction paradigm and quadrant remain not applicable; a single detector does not establish the taxonomy's One Model category. architecturenetworkattributes-future
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Why geometrically correct projection creates a second augmentation problem
Imagine rotating an input image during training. A detector should still place the same car at the same location in the ground plane. Equations (1)–(2) accomplish this by undoing the image augmentation before applying inverse camera intrinsics at each candidate depth. That makes heavy image augmentation compatible with calibrated projection, but it also preserves the spatial layout seen after projection. The paper therefore applies another transformation directly to BEV features and transforms the target boxes with it. The two operations serve different parts of the network. Table 4 makes the consequence concrete: image augmentation alone has a lower best mAP when the BEV encoder is present, whereas adding it on top of BEV augmentation improves performance. This supports the regularization design; coordinate invariance alone does not prove that image augmentation leaves learned features unchanged. projectionaugmentationaugmentation-ablation
Figure 2. Temporary box scaling makes small-object duplicates visible to overlap-based suppression. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the top row first. The red true-positive prediction and yellow false positives have zero intersection-over-union, so the NMS arrow leaves the duplicates in place. In the lower row, scaling increases box size until the predictions overlap; the label changes from zero IoU to positive IoU. NMS can then remove the redundant yellow boxes, and rescaling restores the retained red prediction's size. The blue box is ground truth for illustration. It is not an input used by NMS at inference. Section 3.3 describes category-dependent scale factors and explicitly excludes barriers. scale-nmsnms-ablation
What it supports. Scale-NMS changes the geometry used to decide which predictions are duplicates. It does not claim to improve the underlying depth estimate. The mechanism is especially relevant to objects whose ground-plane footprint is small relative to the detector's output resolution, as tested by the class-wise results in Table 5.
Where the evidence stops. Positive overlap in this schematic is not a universal suppression guarantee; the NMS threshold still matters. Exact scale factors are omitted and were selected on validation data, so their generalization cannot be inferred from this drawing.
5.2 Separate feature learning, duplicate suppression and evaluation geometry
The BEV encoder and CenterPoint head first generate candidate objects. Scale-NMS operates afterward, changing the overlap calculation used to remove redundant candidates. For a small pedestrian footprint, two nearby boxes can refer to the same object while sharing no area; ordinary IoU suppression then misses their relationship. Temporary enlargement makes that relationship visible, and the surviving box is returned to its original scale. Keep this mechanism distinct from the evaluation metric: nuScenes mAP uses ground-plane center-distance matching, not IoU matching. Thus an overlap-based postprocessor can improve a center-distance-based score without changing the learned representation. The pedestrian and cone gains in Table 5 are consistent with this account. They do not establish improved depth accuracy, and the unexplained barrier result prevents attributing every changed column to category scaling. networkscale-nmsprotocolnms-ablation
Figure 3. An occurrence index turns grouped feature accumulation into assignment followed by a sum. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the five point values, all equal to one, under voxel indices 0, 1, 1, 2 and 3. The second occurrence of voxel 1 receives auxiliary index 1, while first occurrences receive index 0. The Assign arrow places values into columns indexed by voxel and rows indexed by occurrence. Summing along the occurrence direction returns values 1, 2, 1 and 1. The duplicate voxel therefore accumulates two contributions without the sequential cumulative-sum construction. Section 4.3 states that fixed camera calibration permits precomputing these assignments at initialization. poolingvalidation
What it supports. The uncapped construction preserves the grouped sum illustrated by the toy example. The authors report that this implementation reduces Tiny's inference latency from 137 to 64 milliseconds, a 53.3% reduction. The mechanism explains why operation scheduling can matter alongside the GFLOPs counted in Table 2.
Where the evidence stops. The implementation consumes extra memory and caps the auxiliary index at 300, dropping excess points. Unlike the uncapped illustration, that step is approximate; the claimed negligible accuracy effect has no numerical error bound or ablation table.
5.3 Treat resolution and pooling as separate deployment choices
The reported speed depends on both how much information the network processes and how the view transformer accumulates it. Table 6 holds Tiny's images at 704×256 while shrinking BEV cells from 0.8 to 0.4 meters: NDS rises from 39.2% to 41.0%, GFLOPs rise from 215.3 to 438.4, and FPS falls from 15.6 to 10.0. Figure 3 addresses a different issue: replacing sequential accumulation with indexed assignment and summation. This accelerates the implementation but adds memory and, with the practical cap, discards some points. A deployment comparison should therefore record image and BEV resolution, pooling implementation and point loss separately. It should also batch the monocular baselines fairly, since dividing single-image throughput by six is an acknowledged limitation of the paper's timing comparison. resolutionpoolingtiming
5.4 Training and inference
During training
Image transformations are inverted during projection, preserving BEV coordinates. The paper consequently augments the view-transformer output and 3D targets together, using flips, rotations within ±22.5 degrees, and scale factors 0.95–1.05. Image augmentation separately uses flips, scaling, rotation within ±5.4 degrees, and cropping. projectionaugmentationtraining
Training uses CBGS, AdamW, gradient clipping, batch size 64 on eight NVIDIA GeForce RTX 3090 GPUs, and a default 20-epoch schedule. Swin learning rate rises from 0.0002 to 0.001 over the first 40% and falls to zero thereafter. ResNet uses tenfold drops at epochs 17 and 20. The text does not specify a full loss formula, loss weights, frozen modules, or a separate training-stage sequence. networktraining
During inference
Default preprocessing deterministically scales and crops each image, followed by image encoding, calibrated lifting, BEV encoding, the detection head and NMS. Reported speed and computational budgets exclude test-time augmentation; the test submission instead uses a single model with augmentation. networkscale-nmstrainingtimingtest
For fixed inference-time camera intrinsics and extrinsics, voxel and auxiliary indices can be precomputed. Assignment into a voxel-by-auxiliary-index matrix followed by summation replaces cumulative-sum pooling. The implementation caps the auxiliary index at 300, discarding excess points; its negligible accuracy impact is an author claim without quantified error here. pooling
5.5 Implementation flow
- Encode images
ResNet or SwinTransformer backbones extract features; a neck fuses resolutions. Tiny uses SwinTransformer-Tiny and FPN-LSS, which upsamples the 1/32-resolution feature and concatenates it with the 1/16-resolution feature. architecturenetwork
- Lift and pool
A categorical depth prediction weights image features through the outer product in Figure 1. Camera geometry places the lifted features in 3D; pooling along the vertical direction produces BEV features. Depth spans 1–60 meters at intervals of 1.25 times the output-feature resolution. Tiny uses a 0.8-meter BEV grid; Base uses 0.4 meters. architecturenetwork
- Encode BEV and detect
Residual blocks and FPN-LSS further encode the shared BEV map. The unmodified first-stage CenterPoint head predicts object properties; its second refinement stage is absent. No action extraction or control-feedback loop is defined. network
- Suppress duplicates
Scale-NMS temporarily scales box sizes by category, performs ordinary overlap-based NMS, and rescales retained boxes. Enlarging small predictions lets nearby duplicates overlap. The method excludes barriers and searches category factors on validation data; the factors themselves are not listed. scale-nms
6. Experiments & results
BEVDet converts calibrated multi-camera images into a shared bird's-eye-view representation and detects present 3D objects there. Its main contribution is making this familiar modular pipeline train effectively: image-space augmentation and BEV-space augmentation address different parts of the network, while Scale-NMS suppresses small-object duplicates. The experiments establish a useful nuScenes accuracy–speed tradeoff under the reported protocols, with weaker attribute prediction and important timing qualifications. This is scene perception, without an action-conditioned dynamics model or demonstrated driving controller.
6.1 Read the original evidence
Table 2. Tiny favors speed; Base spends much more computation for higher detection accuracy. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the modality column to separate camera and LiDAR systems, then compare the unmarked camera rows. Larger mAP and NDS are better; the five error columns favor smaller values. Section 4.1 on page 9 explains that mAP matches by ground-plane center distance and NDS combines the detection indicators. Read image size, GFLOPs and FPS together: neither parameter count nor operation count alone predicts speed. The retained footnote distinguishes FCOS3D initialization, test-time augmentation and ensembles in the lower block. Those marked rows should not silently replace the simpler configurations when making an accuracy–latency comparison. protocolvalidationtimingtest
What it supports. Tiny reports 31.2% mAP and 39.2% NDS at 15.6 FPS, compared with FCOS3D's 29.5%, 37.2% and 1.7 FPS. Base reaches 39.3% mAP and 47.2% NDS at 1.9 FPS. These are two distinct operating points, not one system simultaneously achieving the highest speed and accuracy.
Where the evidence stops. Monocular FPS was divided by six, and the authors acknowledge batching could improve it. The separate test result on page 11 uses train-plus-validation training and augmentation; its 42.2% mAP and 48.2% NDS have no paired FPS here.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Multi-camera 3D detection on nuScenes validation Official 700/150/150 scene split; six cameras, ten classes. Table 2 unmarked camera rows, without test-time augmentation or ensemble. | Tiny: 31.2% mAP, 39.2% NDS, 215.3 GFLOPs, 15.6 FPS at 704×256. Base: 39.3%, 47.2%, 2,962.6 GFLOPs, 1.9 FPS at 1600×640. mAP and NDS (higher better); GFLOPs and FPS | FCOS3D: 29.5% mAP, 37.2% NDS, 2,008.2 GFLOPs, 1.7 FPS at 1600×900. PGD: 33.5% mAP, 40.9% NDS, 1.4 FPS. Tiny improves FCOS3D by 1.7 mAP and 2.0 NDS percentage points. Comparisons use different resolutions and backbones; monocular FPS is divided by six, with acknowledged batching headroom. They do not isolate architecture alone. protocolvalidationtiming |
| Multi-camera 3D detection on nuScenes test BEVDet-Base trained on train plus validation; single model with test-time augmentation; Table 3. | 42.2% mAP; 48.2% NDS mAP; NDS | PGD: 38.6% mAP; 44.8% NDS. The reported gains are 3.6 and 3.4 percentage points. This protocol differs from validation and supplies no corresponding augmented-inference FPS. The historical leaderboard claim is not a present-day ranking. test |
| Augmentation and BEV encoder ablation Tiny on nuScenes validation, fixed 20 epochs; Table 4. | Encoder present: no augmentation 23.0% best/17.4% final; image-only 20.5%/17.8%; BEV-only 26.2%/23.6%; both 31.6%/31.2%. Best-checkpoint and final-epoch mAP | Both augmentations without the BEV encoder: 29.9% best and final. Combined augmentation reduces the peak-to-final drop from 5.6 to 0.4 points. The encoder adds 1.7 points at best checkpoints but 1.3 at the final epoch. Row C's peak epoch is inconsistent between table and prose. augmentation-ablationablation-discrepancies |
| NMS ablation on nuScenes validation BEVDet-Tiny; Tables 5 and Section 4.3. | Scale-NMS 31.2% mAP; pedestrian 34.5% AP; traffic cone 50.0% AP. mAP; class AP | NMS 29.5%, 29.7%, 42.5%, respectively; Circular-NMS 29.8% mAP. Gains concentrate in small objects: pedestrian +4.8 and cone +7.5 points. Validation-tuned factors and an unexplained barrier-row change limit attribution. scale-nmsnms-ablation |
| BEV resolution ablation on nuScenes validation Table 6 rows A and C; Tiny, fixed 704×256 images, BEV cell size 0.8 versus 0.4 meters. | Finer grid: 31.5% mAP, 41.0% NDS, 438.4 GFLOPs, 10.0 FPS. mAP; NDS; GFLOPs; FPS | Coarser grid: 31.2%, 39.2%, 215.3 GFLOPs, 15.6 FPS. Finer BEV resolution improves NDS more than mAP while increasing computation and latency. Resolution is an independent design choice from image size. resolution |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 4. The effect of image augmentation depends on whether BEV augmentation and the BEV encoder are present. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Decode the switches first: IDA is image-view-space data augmentation, BDA is BEV-space data augmentation, and BE is the BEV encoder. Rows A–D retain the encoder; E–H remove it. Compare A with B to add image augmentation alone, A with C to add BEV augmentation alone, and C with D to add image augmentation when BEV augmentation is already active. The best columns include the selected epoch in parentheses; the later mAP and NDS columns report epoch 20. Their parenthetical drops are absolute score differences expressed on the percentage scale, not relative percentage losses. augmentation-ablationablation-discrepancies
What it supports. With BE present, combined augmentation raises best/final mAP from 23.0%/17.4% to 31.6%/31.2%, greatly reducing late-training degradation. Removing BE under combined augmentation gives 29.9%/29.9%. Its contribution is therefore 1.7 points at best checkpoints and 1.3 points at the final epoch.
Where the evidence stops. Row C marks its mAP peak at epoch 11; page 12 says epoch 15. This remains unresolved. The best mAP and best NDS can come from different checkpoints, and no seed variability is reported.
Table 5. Class-wise scores test whether suppression gains concentrate in small-footprint objects. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read downward within a column to hold the object class fixed while changing the suppression strategy. The first numerical column is the mean AP; the remaining columns are class AP, all on the zero-to-one scale. Pedestrian and Traffic Cone are the most direct checks of the small-footprint explanation in Figure 2. Car, Motorcycle and Bicycle provide useful counterexamples to a blanket improvement claim, because their Scale-NMS scores equal the ordinary-NMS scores. Circular-NMS is a separate baseline, so its gains and losses should not be attributed to the proposed scaling mechanism. nms-ablationscale-nms
What it supports. Scale-NMS raises overall mAP from 29.5% to 31.2%. Pedestrian AP rises from 29.7% to 34.5%, and traffic-cone AP from 42.5% to 50.0%. This distribution of gains is consistent with the proposed duplicate-suppression explanation, although it is not a direct count of removed false positives.
Where the evidence stops. Barrier AP changes from 0.467 to 0.498 even though Section 3.3 excludes barriers from scaling. The source does not explain this change, so it cannot be credited to barrier scaling. Factors were tuned on validation data.
7. Analysis & limitations
7.1 What the evidence leaves open
Attribute prediction remains weaker: Base's validation mAAE is 0.191 versus FCOS3D's 0.170, where lower is better. The proposed explanation—that appearance cues are easier to retain in image view—is a conjecture. Joint mapping/detection and driving execution are not evaluated; multi-task learning remains future work. validationattributes-future
Table 4 gives row C's best mAP at epoch 11, whereas page 12 says epoch 15. Table 5 changes barrier AP from 0.467 to 0.498 despite the method's barrier exclusion. Neither discrepancy is resolved by the supplied text. Tables provide no run-to-run uncertainty. ablation-discrepanciesscale-nmsnms-ablationvalidation
7.2 Questions for discussion
- Would the augmentation interaction persist with longer training and matched checkpoint-selection rules?
- How much of the timing advantage remains when monocular baselines batch all six views?
- Does BEV compression discard appearance evidence needed for attributes, and what controlled comparison would isolate that effect?
8. Reproducibility audit
8.1 Requirements and known gaps
A reproduction requires calibrated nuScenes inputs, the exact Table 1 configuration, CBGS, geometric preprocessing and an MMDetection3D implementation. The paper supplies augmentation ranges and training hardware but omits software versions, exact loss configuration, gradient-clip threshold, and numerical Scale-NMS factors. Table 7 also changes module configurations across encoder variants, so it is not a backbone-only swap. networkprotocoltrainingscale-nmsbackbone
Proposed checks: repeat the augmentation factorial experiment with multiple seeds and consistent checkpoint selection; independently compare capped and uncapped auxiliary-index pooling for feature agreement, point loss, latency and detection accuracy. These target the central regularization claim and the unquantified approximation behind the reported pooling speedup. augmentation-ablationpooling
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Repeat the augmentation interaction with a fixed checkpoint rule
Reader-proposed, not run: reproduce the eight IDA/BDA/BE configurations of Table 4 using Tiny, the same nuScenes split, CBGS and 20-epoch schedule. Hold all other settings fixed and repeat with at least three seeds. Report epoch-20 mAP/NDS separately from independently selected best mAP and NDS, retaining the learning curves. The prediction to test is that BDA makes IDA more beneficial when BE is present and reduces the best-to-final drop. If this interaction reverses across seeds or disappears under the fixed final-epoch comparison, the claimed regularization mechanism is less robust than the single reported table suggests. networktrainingaugmentation-ablationablation-discrepancies
Check 2: Measure what the pooling cap trades for speed
Reader-proposed, not run: use one trained Tiny checkpoint and identical calibrated inputs to compare cumulative-sum pooling, uncapped auxiliary-index pooling, and the stated cap of 300. First check uncapped BEV-feature agreement within a declared numerical tolerance. Then measure discarded-point counts, affected-voxel counts, peak memory, synchronized end-to-end latency, mAP and NDS for the capped version, with warmup, precision and hardware held constant. The uncapped versions should agree up to numerical effects; the cap's negligible-impact claim predicts little detection degradation even in samples with crowded voxels. A systematic loss on those samples would falsify that practical claim without disputing the uncapped summation identity. networkprotocolpooling
8.3 Reading coverage
Visual audit: Visually inspected the title, authors, affiliation and v3 date on page 1; architecture on page 2; network descriptions and Table 1 on pages 5–6; both projection equations on page 7; Scale-NMS and dataset setup on page 8; training hardware, schedules, preprocessing and metrics on page 9; benchmark tables and timing qualifications on pages 10–11; augmentation interpretation and conflicting epoch on page 12; NMS and resolution tables on page 13; encoder table, pooling diagram and cap on page 14; future-work statement on page 15. All three figures and all seven tables were visually read. All six final original crops were inspected for readability, complete labels and faithful content. Table 2 retains its necessary variant footnotes; other captions and body prose are excluded from crops. The remaining PDF pages were read in the complete text pass; no appendix or separate supplement was supplied. Code and experiments remain outside this review.
PDF pages inspected for this edition: 1, 2, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract; 1 Introduction (PDF pp. 1–3)
- 2 Related Works: 2.1 Vision-based 2D Perception, 2.2 Semantic Segmentation in BEV, 2.3 Vision-based 3D Object Detection (pp. 3–5)
- 3 Methodology: 3.1 Network Structure, 3.2 The Customized Data Augmentation Strategy, 3.3 Scale-NMS (pp. 5–8)
- 4 Experiment: 4.1 Experimental Settings, 4.2 Benchmark Results, 4.3 Ablation Studies (pp. 8–14)
- 5 Conclusion (pp. 14–15); References (pp. 15–19)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This supplied acquisition caveat was addressed by inspecting all three figures, all seven tables, and the equations in the PDF.
- Separate supplemental material availability has not been fully verified.
- Identity/version scope: the inspected title page identifies arXiv:2112.11790v3 [cs.CV], 16 June 2022, with Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du, all credited to PhiGent Robotics. The catalog title differs only in capitalization and its authors agree. The catalog submission year is 2021; this report reads the supplied 2022 revision. Earlier versions were not supplied, so changes between revisions are unverified.
- All five supplied text chunks were read individually. No appendix is present in the 19-page PDF. Visual inspection covered PDF pages 1, 2, and 5–15; the remaining pages were read as text.
- The linked code was not inspected, dependencies were not installed, and experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title, author block and arXiv version stamp
Exact observed title; Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye and Dalong Du; PhiGent Robotics; arXiv:2112.11790v3, 16 June 2022.
Go to primary source ↓architecturePDF p. 2, Figure 1 and Section 1
Four-module pipeline, outer product and pooling, camera-model input, and Tiny tensor dimensions; BEV detection is motivated alongside BEV semantic mapping.
Go to primary source ↓networkPDF pp. 5–6, Section 2.3 final paragraph; Section 3.1 and Table 1
No LiDAR reliance for depth supervision; ResNet/Swin and FPN/FPN-LSS configurations; depth 1–60 meters at 1.25 times BEV resolution; residual BEV encoder; unmodified first-stage CenterPoint head without refinement.
Go to primary source ↓projectionPDF p. 7, Section 3.2, Equations (1)–(2)
Inverse intrinsics unproject a homogeneous pixel at depth d; inverse augmentation cancels the image transform and preserves its camera-space coordinate.
Go to primary source ↓augmentationPDF p. 7, Section 3.2, BEV Space Learning with Data Augmentation
The paper motivates separate BEV regularization by fewer fused samples and geometric decoupling, transforming BEV features and detection targets together.
Go to primary source ↓scale-nmsPDF pp. 7–8, Section 3.3 and Figure 2
Small nonoverlapping duplicates escape IoU NMS. Category-specific temporary size scaling restores overlap; barriers are excluded and factors are searched on validation data.
Go to primary source ↓protocolPDF pp. 8–9, Section 4.1, Dataset and Evaluation Metrics
nuScenes: 1,000 scenes, six cameras, 700/150/150 split, ten classes; default 51.2-meter ground-plane ROI and 0.8-meter resolution. mAP uses ground-plane center-distance matching; NDS combines detection indicators.
Go to primary source ↓trainingPDF p. 9, Section 4.1, Training Parameters, Data Processing and Inference Speed
AdamW, clipping, batch 64, eight RTX 3090 GPUs, 20 epochs, backbone-specific learning-rate schedules, CBGS, augmentation ranges, deterministic test crop, and MMDetection3D are specified; complete loss and software-version settings are not.
Go to primary source ↓timingPDF pp. 9–10, Section 4.1, Inference Speed
Speed and compute exclude augmentation. Monocular speeds are divided by six; the authors acknowledge this may underestimate a batched implementation's speed.
Go to primary source ↓validationPDF p. 10, Table 2, unmarked FCOS3D, PGD, BEVDet-Tiny and BEVDet-Base rows; Section 4.2
Tiny reports 0.312 mAP/0.392 NDS, 215.3 GFLOPs/15.6 FPS; Base 0.393/0.472, 2962.6/1.9. FCOS3D is 0.295/0.372, 2008.2/1.7; PGD 0.335/0.409, 1.4 FPS. Base/FCOS3D mAAE is 0.191/0.170. Footnotes identify initialization, augmentation and ensemble variants.
Go to primary source ↓testPDF p. 11, Table 3, BEVDet and PGD rows; Section 4.2, nuScenes test set
Train-plus-validation Base with single-model test augmentation scores 0.422 mAP/0.482 NDS; PGD scores 0.386/0.448.
Go to primary source ↓augmentation-ablationPDF p. 11, Table 4, rows A–H and best/final columns; p. 12, Section 4.3, Data Augmentation
Twenty-epoch Tiny factorial ablation toggles IDA, BDA and BE. A–D best/final mAP: .230/.174, .205/.178, .262/.236, .316/.312. H is .299/.299; F image-only without BE is .276/.269.
Go to primary source ↓ablation-discrepanciesPDF p. 11, Table 4, rows C, D and H; p. 12, Data Augmentation, third and fifth paragraphs
Row C is marked e11 in the table but described as epoch 15 in prose. D minus H is .017 at best checkpoints and .013 at the final epoch; prose gives +1.7% without that distinction.
Go to primary source ↓nms-ablationPDF p. 13, Table 5; p. 12, Section 4.3, Scale-NMS
NMS/Circular-NMS/Scale-NMS mAP is .295/.298/.312. NMS to Scale-NMS pedestrian AP is .297 to .345 and cone AP .425 to .500. Barrier AP changes .467 to .498 despite the method's exclusion.
Go to primary source ↓resolutionPDF p. 13, Table 6, rows A and C; Section 4.3, Resolution
At 704×256 input, 0.8 versus 0.4 meter grids yield .312/.315 mAP, .392/.410 NDS, 215.3/438.4 GFLOPs and 15.6/10.0 FPS.
Go to primary source ↓backbonePDF p. 6, Table 1; p. 14, Table 7
Encoder ablations compare complete R50, R101 and Tiny configurations, whose neck, channel counts and BEV block choices differ in Table 1.
Go to primary source ↓poolingPDF p. 14, Figure 3 and Section 4.3, Acceleration
Auxiliary occurrence indices allow assignment followed by summation, with indices precomputed for fixed calibration. Tiny latency drops 137 to 64 ms (53.3%); extra memory and an index cap of 300 are required. Discarded points are claimed to have negligible accuracy impact without quantified results.
Go to primary source ↓attributes-futurePDF p. 11, Section 4.2, attribute-error discussion; p. 15, Section 5
The authors conjecture attribute weakness reflects lost access to appearance cues; improving attributes and studying multi-task learning are future directions.
Go to primary source ↓8.5 Primary sources
BEVDet: High-Performance Multi-Camera 3D Object Detection in Bird-Eye-View ↗
PDF · 8,748 extracted words
Source fingerprint
a54c2013f7d991e1af80d328dc1f3484613bf1197c186b8fb474fd14d61a522b