SliceSemOcc: Vertical Slice 3D Occupancy Framework

• The paper introduces SliceSemOcc, a multimodal framework that utilizes vertical slicing to improve 3D semantic occupancy prediction by preserving height-dependent features.
• It fuses surround-view images and LiDAR data through dual branches, using SEAttention3D and a cross-attention fusion module to refine voxel features.
• Empirical results on nuScenes benchmarks show enhanced mIoU, especially for small-object classes, underscoring the framework’s effectiveness in height-sensitive occupancy prediction.

Searching arXiv for the named method and closely related slice-based occupancy papers. SliceSemOcc is a multimodal 3D semantic occupancy prediction framework for autonomous driving that explicitly models voxel features through vertical slices along the height axis. Introduced in "SliceSemOcc: Vertical Slice Based Multimodal 3D Semantic Occupancy Representation" (Huang et al., 4 Sep 2025), it takes surround-view images and LiDAR sweeps as input and predicts a 3D voxel grid in which each voxel is assigned either empty or one of the semantic classes, with output labels in $[0,16]$ and class $0$ denoting empty space. Its central premise is that even when a perception stack already operates on a 3D voxel grid, height-axis information can still be underused during feature processing, especially when conventional SENet-style channel attention averages away the height dimension (Huang et al., 4 Sep 2025).

1. Task formulation and motivation

SliceSemOcc is formulated as

$V = F\left(I^1, I^2, \cdots, I^N, L\right),$

where $I=\{I^1,\dots,I^N\}$ are the $N$ camera images, $L$ is the LiDAR input, $F$ is the multimodal occupancy prediction network, and $V$ is the semantic occupancy volume (Huang et al., 4 Sep 2025). The task differs from BEV-only scene modeling because occupancy prediction preserves a regular voxel grid over $(X,Y,Z)$, thereby retaining vertical semantic variation and hierarchical structure along height.

The method is motivated by two limitations identified in prior occupancy pipelines. First, height-axis information is often overlooked in voxel feature processing even when the representation itself is three-dimensional. Second, vanilla squeeze-and-excitation style attention globally pools over all spatial dimensions, so all height layers share the same channel weights; the resulting channel reweighting is therefore height-invariant, which SliceSemOcc treats as too coarse for occupancy reasoning (Huang et al., 4 Sep 2025).

The paper grounds this motivation in the empirical observation that, on nuScenes-SurroundOcc, different classes occupy different height ranges. Small objects such as pedestrians and barriers cluster around middle and lower height bands, whereas larger objects such as trucks and buildings span much broader vertical extents. This makes height-sensitive feature refinement particularly relevant for semantic occupancy, especially under complex occlusions and multi-level structures.

2. Multimodal architecture and voxel-space alignment

The overall pipeline is two-branch. In the camera branch, a 2D backbone extracts multi-view image features and a view-to-voxel projection module transforms them into

$$F_C \in \mathbb{R}^{B \times C \times X \times Y \times Z}.$$

In the LiDAR branch, raw points are voxelized and processed by a 3D backbone into

$0$0

After branch-specific vertical-slice refinement, the camera and LiDAR voxel features are concatenated and decoded for occupancy prediction (Huang et al., 4 Sep 2025).

The specified components are a ResNet-50 or ResNet-101 with FPN for the camera backbone, the projection method from InverseMatrixVT3D for image-to-voxel lifting, and VoxelNet for the LiDAR branch. The input LiDAR signal consists of ten LiDAR sweeps. The decoder is described only as a 3D convolutional decoding head that predicts occupancy probabilities and semantic voxel labels; precise decoder architecture details such as number of layers, kernel sizes, intermediate channel counts, or output head decomposition are not provided in the paper (Huang et al., 4 Sep 2025).

A defining architectural choice is voxel-space alignment. Camera features are projected into the same 3D voxel grid as the LiDAR features, LiDAR is voxelized directly into that grid, and fusion is then performed by channel concatenation after branch-specific refinement. The paper does not describe any additional learned deformable alignment or calibration refinement at the multimodal fusion stage.

3. Vertical slice representation, fusion, and SEAttention3D

SliceSemOcc defines its height-aware processing through one global slice and several local slices along the height axis. For nuScenes-SurroundOcc, the total height range is

$0$1

with one global interval

$0$2

and six local intervals

$0$3

The paper states that voxel features are divided along the height-axis into global slices covering the entire range and local slices focusing on key height bands, but it does not provide a formal equation for the actual slice extraction operator. Accordingly, the exact slice construction implementation is unspecified beyond the conceptual partitioning of the height axis (Huang et al., 4 Sep 2025).

The global slice is intended to capture holistic scene structure, long-range vertical dependencies, and large-object extent, producing a global voxel feature $0$4. The local slices are intended to focus on key vertical bands where small objects commonly occur, producing a local voxel feature $0$5. The paper states that for each global and local slice, it independently applies SEAttention3D, and then integrates the resulting slice features via a 3D convolutional module to obtain $0$6 and $0$7. Kernel sizes, layer counts, and the exact aggregation rule across multiple local slices are not fully specified (Huang et al., 4 Sep 2025).

The Vertical Slice Fusion module then performs a bidirectional cross-attention mechanism, although the described operation is mathematically closer to mutual spatial reweighting than to full query-key-value transformer cross-attention. Each branch is first reduced from $0$8 channels to a single-channel 3D attention map by a 3D convolution, yielding $0$9 and $V = F\left(I^1, I^2, \cdots, I^N, L\right),$0. The reweighting is

$V = F\left(I^1, I^2, \cdots, I^N, L\right),$1

Afterward, the recalibrated features are concatenated along channels, reduced back to $V = F\left(I^1, I^2, \cdots, I^N, L\right),$2 channels by a 3D convolution, and passed through SEAttention3D again to produce $V = F\left(I^1, I^2, \cdots, I^N, L\right),$3 (Huang et al., 4 Sep 2025).

SEAttention3D is the paper’s height-preserving variant of squeeze-and-excitation. Given

$V = F\left(I^1, I^2, \cdots, I^N, L\right),$4

it averages only over the planar dimensions: $V = F\left(I^1, I^2, \cdots, I^N, L\right),$5 yielding

$V = F\left(I^1, I^2, \cdots, I^N, L\right),$6

This height-wise descriptor is processed by two 1D convolutions with a ReLU and sigmoid, producing $V = F\left(I^1, I^2, \cdots, I^N, L\right),$7, which is then broadcast back: $V = F\left(I^1, I^2, \cdots, I^N, L\right),$8 The final reweighting is

$V = F\left(I^1, I^2, \cdots, I^N, L\right),$9

Unlike vanilla SENet, which would produce one weight per channel shared across all heights, SEAttention3D produces per-channel, per-height weights and therefore preserves height-wise variation during attention (Huang et al., 4 Sep 2025).

4. Supervision, datasets, and evaluation protocol

The total loss is stated as the sum of focal loss, Lovász-Softmax loss, geometric affinity loss, and semantic affinity loss: $I=\{I^1,\dots,I^N\}$0 The notation appears corrupted in the manuscript, but the intended meaning is clear from the accompanying text and the MonoScene reference. The exact formulas for the focal, Lovász, and affinity terms are not reproduced in the paper (Huang et al., 4 Sep 2025).

Evaluation is performed on two nuScenes-derived occupancy benchmarks. For nuScenes-SurroundOcc, the occupancy annotation range is $I=\{I^1,\dots,I^N\}$1 m and $I=\{I^1,\dots,I^N\}$2 m with grid resolution

$I=\{I^1,\dots,I^N\}$3

For nuScenes-OpenOccupancy, the evaluation range is $I=\{I^1,\dots,I^N\}$4 m and $I=\{I^1,\dots,I^N\}$5 m, with grid dimensions reported as

$I=\{I^1,\dots,I^N\}$6

The paper notes that axis ordering for this latter grid is not clearly explained, so only the reported grid dimensions can be treated as specified (Huang et al., 4 Sep 2025).

The main metric is mean IoU (mIoU) across semantic classes. Tables also report class-wise IoUs for 16 occupied classes, including barrier, bicycle, bus, car, construction vehicle, motorcycle, pedestrian, traffic cone, trailer, truck, drivable surface, other flat, sidewalk, terrain, manmade, and vegetation. Beyond backbone and modality choices, several reproduction-critical details remain unspecified, including batch size, optimizer, learning rate, schedule, augmentation, the reduction ratio $I=\{I^1,\dots,I^N\}$7 in SEAttention3D, exact decoder structure, and training hardware (Huang et al., 4 Sep 2025).

5. Quantitative results and ablation findings

On nuScenes-SurroundOcc, SliceSemOcc achieves

$I=\{I^1,\dots,I^N\}$8

exceeding M-CONet at $I=\{I^1,\dots,I^N\}$9, Co-Occ at $N$0, and OccFusion at $N$1 (Huang et al., 4 Sep 2025). Compared with OccFusion, the absolute gain is $N$2 mIoU; compared with M-CONet, the gain is $N$3 mIoU, which the paper describes as a 14.2% relative gain. Selected class-wise IoUs on SurroundOcc for SliceSemOcc are 30.1 for barrier, 19.9 for bicycle, 36.5 for bus, 38.9 for car, 24.7 for motorcycle, 26.0 for pedestrian, 17.7 for traffic cone, 32.2 for truck, and 40.8 for vegetation (Huang et al., 4 Sep 2025).

On nuScenes-OpenOccupancy, SliceSemOcc achieves

$N$4

matching the listed OccLoff mIoU of $N$5 and exceeding M-CONet at $N$6 and Co-Occ at $N$7 (Huang et al., 4 Sep 2025). The paper highlights especially strong gains over M-CONet for small-object classes: bicycle $N$8 vs $N$9, motorcycle $L$0 vs $L$1, pedestrian $L$2 vs $L$3, and trailer $L$4 vs $L$5.

A major empirical claim is that SliceSemOcc is particularly effective on small-object categories. On SurroundOcc, the paper explicitly reports relative gains over M-CONet of +21.3% for barrier, +53.1% for bicycle, +37.2% for motorcycle, +30.0% for pedestrian, and +20.4% for traffic cone. This is consistent with the design choice of local height slices focused on the vertical bands where small objects tend to occur (Huang et al., 4 Sep 2025).

The ablation studies decompose the contribution of the method’s main components. With neither local nor global slices, mIoU is 26.98 at 5.6 GB memory; local slices alone yield 27.24 at 5.7 GB; global slices alone yield 27.41 at 5.7 GB; and both together yield 28.21 at 6.2 GB. Three local slicing schemes are compared: fine uniform bins $L$6 give 27.67 mIoU, coarse uniform bins $L$7 give 27.75, and the dataset-inspired nonuniform bins $L$8 give 28.21 (Huang et al., 4 Sep 2025).

The paper also isolates the effect of SEAttention3D and the fusion rule. Replacing vanilla SENet with SEAttention3D improves mIoU from 28.08% to 28.21%, a gain of +0.13 mIoU. Replacing simple concatenation fusion (27.97) with the global-local cross-attention fusion yields 28.21, a gain of +0.24 mIoU. The full Vertical Slice Fusion module adds about 0.6 GB of memory relative to the no-slice version. No runtime or FPS numbers are reported (Huang et al., 4 Sep 2025).

6. Interpretation, limitations, and relation to adjacent slice-based work

SliceSemOcc’s significance lies in a narrow but consequential claim: in 3D semantic occupancy, using a voxel grid is not by itself sufficient to exploit the third dimension effectively. The framework operationalizes height-aware reasoning through dual-scale height slicing, bidirectional global-local reweighting, and a squeeze-and-excitation variant that preserves height resolution. The empirical pattern suggests that these mechanisms are most useful when classes occupy distinct vertical bands and when small objects must be distinguished under clutter or occlusion (Huang et al., 4 Sep 2025).

The method also has explicit and implied limitations. Performance is weaker on drivable surface, other flat, sidewalk, and terrain, which the authors attribute to similar geometry and sensor appearance across these flat ground-like categories; even with height-aware attention, adjacent height layers may still receive similar weights. Additional limitations include the extra memory cost from slice processing, the fact that several implementation details are not fully specified, the modest standalone gain from SEAttention3D, and the dataset-tuned nature of the best local height intervals. The authors explicitly identify future work as making SliceSemOcc more lightweight to reduce training and validation computation (Huang et al., 4 Sep 2025).

A common terminological misconception is to equate SliceSemOcc with all other "slice"-based methods. The closest architectural relative in the provided literature is SliceOcc, which is an RGB camera-based model specifically tailored for indoor 3D semantic occupancy prediction and uses a vertical slice representation with pairs of floor and ceiling planes; on EmbodiedScan it reports 15.45% mIoU over 81 indoor categories (Li et al., 28 Jan 2025). The overlap is conceptual rather than identity: SliceOcc addresses indoor RGB occupancy with planar slice queries, whereas SliceSemOcc is a camera-LiDAR multimodal autonomous-driving framework that slices voxel features along height and adds height-preserving channel attention. By contrast, SemSlicer is a framework for semantic data slicing using LLMs and is explicitly not about occupancy, 3D scene understanding, or segmentation (Yang et al., 2024), while NeuSemSlice concerns neuron-level semantic slicing for DNN maintenance rather than scene occupancy (Zhou et al., 2024). This suggests that "slice" in SliceSemOcc should be understood specifically as height-axis feature partitioning within a voxel occupancy model, not as a generic reference to semantic slicing in other areas.

Within the trajectory of slice-based occupancy research, SliceSemOcc can therefore be understood as a height-aware voxel feature refinement framework rather than a wholly new end-to-end occupancy paradigm. Its contribution is to make semantic occupancy prediction use the height axis more explicitly and more selectively, with the strongest reported gains appearing on categories whose semantics are tightly coupled to particular height bands (Huang et al., 4 Sep 2025).