IndoorBEV: Mask-Centric BEV Perception

• IndoorBEV is a mask-centric BEV perception framework that produces class-labeled object detections and dense 2D footprint masks from a single LiDAR scan.
• It employs an axis-fusion encoder and SwinTransformer backbone to capture fine-grained, occlusion-completed footprints for improved navigation and collision checking.
• The method leverages holistic BEV masks to handle irregular indoor geometries and sparse point clouds, making it effective in cluttered indoor environments.

Searching arXiv for the specified paper to ground the response in the current record. IndoorBEV is a mask-centric Bird’s-Eye View (BEV) perception framework for indoor mobile robots that operates on a single LiDAR point cloud and produces class-labeled object detections, 3D attributes, and dense 2D footprint masks in the ground plane, including completion of missing contours caused by occlusion or sparsity (Li et al., 23 Jul 2025). It is formulated for cluttered indoor scenes containing both static objects such as walls, furniture, and shelves and dynamic entities such as people, robots, and moving items. The central premise is that indoor robotics benefits from representing scene occupancy directly in BEV mask space rather than relying exclusively on oriented 3D cuboids, because downstream navigation, collision checking, and planning typically consume top-down occupancy or cost maps rather than volumetric box abstractions.

1. Problem formulation and representational rationale

IndoorBEV targets a specific perception problem: from a single LiDAR scan, generate a BEV map in which relevant indoor objects are detected and classified, localized in 3D by position, yaw, and dimensions, and represented by dense 2D footprint masks on the ground plane (Li et al., 23 Jul 2025). The formulation is motivated by several properties of indoor environments stated explicitly in the paper: high clutter, diverse and irregular geometries, dense co-location of static and dynamic objects, heavy occlusions, and sparse or partial returns from compact sensors such as the Livox MID-360.

The method is positioned against traditional 3D detection pipelines that regress oriented 3D bounding boxes from point clouds. In the paper’s framing, such pipelines struggle to capture the fine-grained footprint needed for safe navigation, handle non-rectangular shapes, and provide outputs that are directly usable by navigation stacks. By contrast, a BEV representation provides a top-down, metric-aligned map aligned with the robot’s planning frame, collapses vertical-axis occlusions into ground-plane occupancy, and plugs directly into costmap-based navigation, collision checking, and path planning.

A recurrent misconception in indoor perception is that a BEV representation necessarily discards information needed for object-level reasoning. IndoorBEV does not adopt that reduction. It predicts dense BEV masks while still supervising dimensions, position, and yaw, so the mask is the primary occupancy representation rather than a substitute for all geometric attributes. The system therefore couples instance segmentation in BEV with auxiliary 3D localization.

2. BEV discretization and axis-fusion encoding

IndoorBEV begins from a LiDAR point cloud

$$P = \{p_i\}_{i=1}^N,\qquad p_i = (x_i, y_i, z_i, I_i),$$

expressed in the LiDAR or robot frame (Li et al., 23 Jul 2025). The perception workspace is a fixed cuboid,

$$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$

which is discretized into 3D clusters of uniform size

$v = 0.02\ \text{m}.$

This yields

$$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$

so the BEV grid is $500 \times 500$, with each cell representing a $2$ cm by $2$ cm footprint.

Rather than padding clusters and applying PointNet independently per voxel, the method uses an Axis-Fusion Point Cloud Compact Network (AF-PCCN). The encoder groups points into 3D clusters based on $(x,y,z)$, learns X–Z and Y–Z axis feature views, and fuses these axis-specific features into a single embedding of dimension $F=C_{enc}$, with $C_{enc}=64$ in the implementation. Empty clusters are treated as noise rather than padded. The scattered BEV pseudo-image is

$$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$0

For a point $$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$1, the corresponding BEV indices are

$$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$2

Features from points falling into the cluster at $$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$3 are aggregated and fused along the vertical axis and the axis-specific views to obtain a BEV cell feature vector. The exact aggregation is handled by AF-PCCN, and the paper emphasizes that height information is encoded implicitly in learned cluster features rather than through explicit height channels. After LayerNorm, the encoder output is the initial BEV feature map

$$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$4

The stated design rationale is tailored to indoor point clouds: they are highly sparse, locally anisotropic, and often populated by non-empty clusters containing very few points. Axis fusion is therefore used to preserve vertical structure compactly while reducing the overhead associated with padding empty space.

3. Backbone and query-based prediction head

The full architecture consists of three stages: the AF-PCCN axis compact encoder, a window-based BEV backbone implemented as a Custom SwinTransformer, and a query-based decoder head following DETR/Mask2Former-style principles (Li et al., 23 Jul 2025).

The Custom SwinTransformer processes the BEV tensor in four stages with strides $$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$5, depths $$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$6, attention heads $$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$7, window size $$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$8, MLP ratio $$[x_{\min}, x_{\max}] \times [y_{\min}, y_{\max}] \times [z_{\min}, z_{\max}] = [-5,5] \times [-5,5] \times [-3,3]\ \text{meters},$$9, and gradient checkpointing enabled. Stage 1 retains the full $v = 0.02\ \text{m}.$0 spatial resolution, while later stages downsample by factors of $v = 0.02\ \text{m}.$1. Each block alternates between Windowed Multi-Head Self-Attention and Shifted-Window Multi-Head Self-Attention. For a window feature matrix $v = 0.02\ \text{m}.$2, attention is written as

$v = 0.02\ \text{m}.$3

The paper’s interpretation is that shifted windows allow global spatial context to emerge despite local attention windows. The final-stage feature map

$v = 0.02\ \text{m}.$4

is then upsampled back to the original BEV resolution.

The decoder first applies an adaptive upsampler and a $v = 0.02\ \text{m}.$5 convolution plus normalization so that

$v = 0.02\ \text{m}.$6

This projected BEV map serves as decoder memory. IndoorBEV uses $v = 0.02\ \text{m}.$7 learnable object queries,

$v = 0.02\ \text{m}.$8

processed by four transformer decoder layers. Each layer applies self-attention among queries, cross-attention against the flattened BEV sequence, and an FFN with normalization. No NMS is used; one-to-one assignments are handled through DETR-style set prediction and Hungarian matching.

From each refined query embedding, the decoder predicts class logits, 3D attributes, and a mask embedding. The 3D attributes are dimensions $v = 0.02\ \text{m}.$9, position $$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$0, and yaw. The paper states that the class logits include a background (“no-object”) class and separately states, in the training details, that the system uses 9 foreground classes plus background; this suggests that the experimental setup should be read in terms of foreground categories plus an explicit background label. Mask logits are computed by inner products between a query-specific mask embedding and BEV features:

$$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$1

After sigmoid activation, $$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$2 is interpreted as the probability that cell $$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$3 belongs to object $$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$4.

4. Mask-centric prediction and footprint completion

IndoorBEV reformulates detection as set-based instance segmentation in BEV rather than as box regression followed by analytic footprint derivation (Li et al., 23 Jul 2025). Each object query predicts a dense mask

$$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$5

representing the object footprint directly on the ground plane. The stated advantages in indoor scenarios are the ability to handle arbitrary shapes and sizes, encode geometry that is directly relevant to navigation, and avoid brittle box-fitting procedures on sparse or partially observed LiDAR returns.

A central contribution is footprint completion. The paper explicitly contrasts raw BEV point footprints with completed masks and argues that the latter emerge from three mechanisms: global context from the Swin backbone, learned instance priors encoded in the object queries, and mask decoding from contextualized BEV features. Because training supervises complete ground-truth footprints rather than only visible points, the mask is not constrained to be non-zero only where LiDAR returns exist. Formally, if

$$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$6

then positive logits may still be produced in cells lacking direct returns when neighboring context and class priors support a completed footprint.

This choice is especially consequential for indoor clutter. Furniture with legs, open structures, partial self-occlusion, or complex contours is difficult to approximate with a cuboid. The mask-centric formulation instead represents occupancy at the granularity required for top-down planning. The same mechanism is used for static and dynamic objects. Differentiation occurs only through class labels and semantics; the architecture and loss formulation treat them uniformly.

5. Supervision, dataset construction, and empirical results

Training follows DETR-style set prediction with Hungarian matching (Li et al., 23 Jul 2025). For a predicted query $$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$7 and ground-truth object $$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$8, the matching cost combines classification, mask, and regression terms:

$$H = \frac{y_{\max}-y_{\min}}{v} = \frac{10}{0.02} = 500,\qquad W = \frac{x_{\max}-x_{\min}}{v} = \frac{10}{0.02} = 500,$$9

Once matching is fixed, the loss is

$500 \times 500$0

The paper gives the weights

$500 \times 500$1

Auxiliary losses are applied at intermediate decoder layers, and unmatched queries are trained toward the background class with weight $500 \times 500$2.

Evaluation uses a custom hybrid dataset combining simulated and real-world LiDAR. Simulated data is generated in MuJoCo with procedural indoor layouts and objects such as planes, boxes, spheres, cylinders, capsules, and ellipsoids. LiDAR simulation uses Taichi-based ray casting with ray direction

$500 \times 500$3

The simulated pipeline provides perfect ground-truth annotations for class, pose, dimensions, and footprint masks. Real-world data uses a Livox MID-360 mounted on a mobile robot in offices, labs, and corridors, recorded via ROS2 and stored as .bin $500 \times 500$4 arrays. Annotation follows KITTI label_2 style, and footprint masks are produced by projecting 3D boxes to BEV and rasterizing polygons with MaskRasterizer. The semi-automatic annotation pipeline uses SUSTechPOINTS tools with manual verification. The real-world split is approximately 4.5k training frames, 1k validation frames, and 1k test frames.

Training details are specified as a $500 \times 500$5 grid over $500 \times 500$6 m, encoder output channel $500 \times 500$7, $500 \times 500$8 queries, hidden dimension $500 \times 500$9, four decoder layers, AdamW with learning rate $2$0, weight decay $2$1, cosine learning-rate schedule with warm-up, batch size $2$2, and approximately 12k iterations on multiple NVIDIA A100 GPUs. Augmentations include random flipping, rotation, global scaling, range dropout, jitter, and object-level perturbations and insertion. At inference, predictions are filtered by class probability threshold $2$3, and no NMS is used.

On the real-world test set, the reported metrics are [email protected] (mask IoU) 78.4%, [email protected] (mask IoU) 63.7%, mIoU 67.2%, and Panoptic Quality 64.5%. These metrics are computed in BEV mask space rather than by 3D box IoU. The paper does not provide detailed numerical baselines for all methods, but it explicitly argues that bounding-box-based BEV detectors struggle with irregular shapes and clutter, whereas IndoorBEV yields more accurate footprints, especially in occluded or partially visible scenarios. It also notes that formal numeric ablations are not detailed, though the axis-fusion encoder, window-based backbone, and query-based mask head are highlighted conceptually as the main contributors.

6. Robotic use, limitations, and prospective extensions

The output representation is designed to be directly usable by downstream robotic systems (Li et al., 23 Jul 2025). BEV masks can be converted into occupancy or cost maps; static classes such as walls and fixed furniture provide persistent obstacles; and the $2$4 cm resolution supports tight-clearance planning in cluttered interiors. For collision avoidance, precise footprints reduce dependence on rough box inflation, and footprint completion is intended to suppress “phantom gaps” that could otherwise induce unsafe trajectories through occluded space. For path planning, the paper states that the BEV map can be integrated with global and local planners such as Nav2 costmaps. Dynamic agents, represented as instance-level masks in BEV, can be tracked over time and supplied to motion prediction modules, with the top-down representation simplifying integration with classical multi-target tracking and 2D velocity estimation.

The authors also identify several limitations. First, the Livox MID-360 produces relatively low-density point clouds, and this sparsity can degrade detection and footprint quality in highly cluttered or distant regions. Second, the real-world dataset is collected from only five environments, which may limit generalization across building types and layout diversity. Third, although the method is designed for real-time use, the Swin-plus-Transformer architecture may be heavy for embedded robotic platforms, so latency–accuracy trade-offs remain unresolved on lower-end hardware.

Future directions listed in the paper include expanded real-world data collection across more environments and sensor setups, broader comparative evaluation against classical LiDAR detectors and BEV segmentation frameworks, optimization for real-time deployment through pruning, quantization, hardware acceleration such as TensorRT and sparse kernels, explicit temporal modeling for tracking and motion prediction, and multi-sensor fusion with cameras to address difficult surfaces such as glass and highly specular materials. A plausible implication is that the framework is best understood not as a generic indoor detector, but as a perception layer specifically aligned with how indoor navigation stacks already represent traversability: top-down, instance-aware, and occupancy-centric.