MapKD: Distillation for Efficient HD Mapping

• The paper introduces a three-stage teacher–coach–student framework that distills multimodal, prior-based supervision into a lightweight camera-only model for HD map construction.
• It employs Token-Guided 2D Patch Distillation and Masked Semantic Response Distillation to align BEV features and semantic outputs effectively.
• Empirical results on nuScenes show marked improvements in mIoU and mAP while maintaining real-time inference speeds without LiDAR or offline maps.

Searching arXiv for the MapKD paper and closely related online HD map construction/distillation work. Calling arXiv search for “MapKD online HD map construction knowledge distillation”. MapKD is a three-stage teacher–coach–student knowledge distillation framework for efficient online HD map construction from camera images. Its central objective is to allow a lightweight camera-only student to absorb the benefits of multimodal perception and map priors without requiring LiDAR, SD maps, or HD maps at inference. In the formulation reported for nuScenes, the task is posed in Bird’s-Eye View (BEV) and targets pedestrian crossings, lane dividers, and road boundaries; the method is motivated by the gap between efficient but weaker camera-only systems and accurate but expensive multimodal systems that depend on stale offline maps and heavier sensor suites (Yan et al., 21 Aug 2025).

1. Problem formulation and motivation

Online HD map construction predicts semantic map elements around the ego vehicle in real time, typically in BEV. The setting emphasized for MapKD centers on three map classes in nuScenes: pedestrian crossings, lane dividers, and road boundaries. The methodological tension is explicit: camera-only models are efficient but struggle with geometry, occlusion, and long-range structure, whereas multimodal models with LiDAR and map priors are accurate but expensive and not always deployable. Offline priors such as SD maps and HD maps can improve accuracy, but they introduce dependence on stale external data and additional preprocessing cost (Yan et al., 21 Aug 2025).

MapKD is designed to retain the deployment simplicity of a vision-centric model while transferring structural information from stronger training-time sources. The key observation is that direct teacher-to-student distillation is difficult because the teacher and student differ simultaneously in modality, representation, and capacity. The teacher has access to LiDAR, cameras, and SD/HD priors; the student sees only cameras; and the teacher’s BEV features are richer and more structured than the student’s. This motivates an intermediate model that narrows the transfer gap.

2. Teacher–Coach–Student architecture

The architectural core of MapKD is the Teacher–Coach–Student (TCS) paradigm. The teacher is a full multimodal model with priors, the coach is a camera-based intermediate model augmented with simulated LiDAR and priors, and the student is a lightweight camera-only model used for deployment (Yan et al., 21 Aug 2025).

Component	Inputs	Role
Teacher $\mathcal{T}$	Multi-view images, real LiDAR, SD map prior, HD map prior	Strongest supervision with rich geometry and priors
Coach $\mathcal{C}$	Images, simulated LiDAR features, SD map prior, HD map prior	Bridge between multimodal teacher and camera-only student
Student $\mathcal{S}$	Images only	Lightweight deployment model

The teacher is based on PMapNet and is defined by

$$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$

$$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$

$$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$

$$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$

The teacher supplies the richest geometric and semantic targets.

The coach uses the same high-level architecture as the teacher, but replaces real LiDAR with simulated LiDAR features generated from images: $$\tilde{F}_L = \mathrm{Enc}_{\mathrm{3D}}(I), \qquad F_I = \mathrm{Enc}_{\mathrm{2D}}(I),$$

$$F_C^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, \tilde{F}_L)),$$

$$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_C^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$

$\mathcal{C}$0

Its stated function is to be modality-aligned with the student while still carrying structural information from priors and pseudo-LiDAR.

The student is HDMapNet-style, lightweight, and camera-only: $\mathcal{C}$1

$\mathcal{C}$2

$\mathcal{C}$3

Only this student is retained at inference. A central point in the paper is that the coach is not an auxiliary detail but the principal mechanism for bridging the cross-modal knowledge transfer gap.

3. Distillation mechanisms

MapKD introduces two targeted distillation strategies: Token-Guided 2D Patch Distillation (TGPD) for BEV feature alignment and Masked Semantic Response Distillation (MSRD) for semantic learning guidance (Yan et al., 21 Aug 2025).

TGPD is the feature-level module. Its goal is to align BEV representations while preserving local spatial details and global scene context. For a BEV feature map $\mathcal{C}$4, the procedure is to divide the feature map into 2D patches, linearly project each patch into a token, compute a global token by average pooling, concatenate the global token and patch tokens, apply attention over the token sequence, and align the student attention and embeddings to those of teacher and coach. The tokenization is summarized as

$\mathcal{C}$5

$\mathcal{C}$6

The teacher-to-student and coach-to-student TGPD losses are given as

$\mathcal{C}$7

$\mathcal{C}$8

with

$\mathcal{C}$9

In the interpretation stated by the paper, the KL term aligns attention distributions and the MSE term aligns BEV token or feature values; the teacher contributes richer targets and the coach contributes a more student-compatible target.

MSRD is the output-level module. Distillation is restricted to meaningful regions by deriving a binary mask $\mathcal{S}$0 from ground truth so that supervision focuses on foreground map elements rather than abundant BEV background. The masked logits are

$\mathcal{S}$1

with teacher and coach probabilities

$\mathcal{S}$2

The loss is

$\mathcal{S}$3

This transfers semantic knowledge at the logit level only on foreground map regions.

4. Objective function, priors, and training protocol

The full training objective combines the base supervised loss with BEV-level and output-level distillation terms (Yan et al., 21 Aug 2025): $\mathcal{S}$4 where

$\mathcal{S}$5

$\mathcal{S}$6

and

$\mathcal{S}$7

Prior knowledge enters through SD maps and HD maps used in the teacher and coach pipelines via $\mathcal{S}$8 and $\mathcal{S}$9. The student does not require these priors at inference. A common simplification of MapKD is to describe it as a prior-based deployment method; that would be inaccurate. The priors are training-time sources of supervision and structural regularization, not deployment-time dependencies.

The reported experiments are conducted on nuScenes under the standard online map construction benchmark setup. Training details provided in the appendix include a BEV range of $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$0, voxel resolution $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$1, input image size $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$2, batch size $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$3, $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$4 GPUs, $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$5 workers, Adam with learning rate $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$6, weight decay $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$7, gradient clipping $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$8, and $$F_I = \mathrm{Enc}_{\mathrm{img}}(I), \qquad F_L = \mathrm{Enc}_{\mathrm{lidar}}(L),$$9 epochs for teacher and coach pretraining; student distillation in the main comparison is reported with $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$0 epochs. The best reported loss weights are $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$1, $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$2, $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$3, and $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$4.

5. Empirical performance and ablation structure

The reported evaluation uses IoU and AP for each of the three classes, together with mIoU, mAP, and FPS. The main baseline student is HDMapNet, with mIoU $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$5, mAP $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$6, and FPS $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$7. MapKD achieves mIoU $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$8, mAP $$F_T^{\mathrm{BEV}} = \mathrm{BEVProj}(\mathrm{Concat}(F_I, F_L)),$$9, and FPS $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$0, corresponding to improvements of $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$1 mIoU and $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$2 mAP while essentially preserving inference speed (Yan et al., 21 Aug 2025).

Model	mIoU	mAP
HDMapNet	31.16	23.13
BEVDistill	32.33	24.99
UniDistill	32.00	25.44
MapDistill	32.77	26.90
MapKD	37.84	34.07

The comparisons are particularly structured around training efficiency and transfer quality. At $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$3 epochs, MapKD surpasses BEVDistill, UniDistill, and MapDistill; at $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$4 epochs those prior methods improve to $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$5, $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$6, and $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$7, respectively, but MapKD remains best with only $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$8 epochs at $$F_{\mathrm{BEV}}^{\mathrm{HD}} = \mathrm{HDFusion}(\mathrm{SDFusion}(F_T^{\mathrm{BEV}}, M_{\mathrm{SD}}), M_{\mathrm{HD}}),$$9. The paper therefore characterizes MapKD as both more accurate and more efficient in training than these alternatives.

Ablation results separate the contributions of the TCS framework and the two distillation losses. Starting from the baseline $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$0, only MSRD yields $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$1, only TGPD yields $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$2, TGPD + MSRD yields $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$3, TCS + MSRD yields $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$4, TCS + TGPD yields $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$5, and full MapKD yields $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$6. The two-stage versus three-stage comparison is also explicit: direct teacher-student distillation with the best teacher and TGPD/MSRD reaches $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$7, whereas three-stage TCS reaches $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$8. This isolates the coach as performance-relevant rather than merely architectural ornamentation.

The effect of priors is reported separately. Without distillation, the plain student achieves $$S_T = \mathrm{Dec}_{\mathrm{sem}}(F_{\mathrm{BEV}}^{\mathrm{HD}}).$$9, the student with SD priors achieves $$\tilde{F}_L = \mathrm{Enc}_{\mathrm{3D}}(I), \qquad F_I = \mathrm{Enc}_{\mathrm{2D}}(I),$$0, and the student with SD + HD priors achieves $$\tilde{F}_L = \mathrm{Enc}_{\mathrm{3D}}(I), \qquad F_I = \mathrm{Enc}_{\mathrm{2D}}(I),$$1. With MapKD, the model uses only camera input at inference yet achieves $$\tilde{F}_L = \mathrm{Enc}_{\mathrm{3D}}(I), \qquad F_I = \mathrm{Enc}_{\mathrm{2D}}(I),$$2. This establishes the intended trade-off: removal of deployment-time prior dependence while retaining competitive or better performance.

6. Interpretation, qualitative behavior, and limitations

The qualitative results are described in terms of BEV structure and map completeness. The raw student produces blurry, incomplete BEV features and map outputs; MapDistill improves these somewhat; MapKD yields more coherent BEV structure and more complete map elements; the coach’s BEV features are structurally better than the student’s; and the teacher remains strongest, though the student under MapKD approaches it more closely. The multi-scene examples suggest better handling of occlusions, intersections, complex road geometries, and small boundary details (Yan et al., 21 Aug 2025).

Several misconceptions are addressed directly by the reported design. First, MapKD is not a multimodal inference system: only the student is deployed. Second, it is not simply a two-loss distillation recipe: the coach is presented as the main novelty beyond standard two-stage distillation because it shares the student’s modality while still being enhanced with pseudo-LiDAR and priors. Third, the method does not eliminate the cost of strong priors and additional modalities altogether; rather, it relocates their use to training.

The paper does not provide an explicit limitations section, but the method still depends on a strong pretrained teacher and coach, access to SD/HD priors during training, and simulated LiDAR generation in the coach. A plausible implication is that the training pipeline remains substantially more complex than a plain student-only system even though inference remains camera-only and fast. In that sense, MapKD is best understood as a deployment-efficiency framework built on comparatively rich supervision during training.

Within the specific context of online HD map construction, MapKD occupies a distinct position: it treats multimodal sensing and map priors as knowledge sources to be distilled rather than as mandatory inference-time inputs. The resulting formulation is therefore a transfer architecture for structural BEV regularities, not merely a conventional compression scheme.