acceptodds
Withdrawn from ICLR 2027

UAVOcc: Hierarchical Spatiotemporal Semantic Occupancy Prediction for UAVs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Aerial semantic occupancy faces a resolution–coverage trade-off under a fixed voxel budget: fine discretization preserves local geometry but restricts physical extent, while coarse discretization expands coverage at the expense of detail. We formulate hierarchical aerial semantic occupancy, representing these competing requirements with complementary fine-local L1 and coarse large-range L2 physical levels of distinct metric resolutions and extents. To make this formulation measurable, we establish a benchmark spanning three virtual and real-world aerial data sources, with a common L1/L2 task definition and Package-Physical evaluation for non-redundant joint physical coverage. Modeling this hierarchy requires cross-level interaction beyond ordinary multi-scale fusion, as the two levels differ in metric correspondence and temporal support. We therefore develop UAVOcc with Global–Local Cross-Level Interaction (GLCI), which combines metric-local 3D correspondence with broader global context through bidirectional spatial exchange; pose-aligned same-level temporal fusion is further followed by coarse-to-fine Temporal GLCI to exploit the broader historical support of L2. Across all three benchmarks, UAVOcc consistently outperforms adapted occupancy baselines, improving Package-Physical mIoU by 3.41, 5.47, and 5.27 points over the strongest listed baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial Semantic Occupancy, UAV Perception, Hierarchical Spatiotemporal Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NfpyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UAVOcc, a hierarchical semantic occupancy framework for UAV perception. The main idea is to represent the scene using two complementary physical levels with different resolutions. To facilitate information exchange across these levels, the authors introduce Global–Local Cross-Level Interaction (GLCI), which combines local cross-attention with global cross-attention. The framework further incorporates temporal fusion and a coarse-to-fine Temporal GLCI module. Experiments across all three benchmarks demonstrate consistent improvements.

Critical strengths

This paper addresses a practically relevant issue in aerial occupancy prediction, where fine spatial resolution and large physical coverage are difficult to achieve simultaneously.

Critical weaknesses

1.The proposed method is not well aligned with the paper’s motivation. This paper motivates hierarchical occupancy through the resolution–coverage trade-off under a fixed voxel budget. However, UAVOcc does not actually resolve this trade-off. Instead, it maintains two complete occupancy representations, with both L1 and L2 containing 96x96x96 voxels. Therefore, this method increases the total representation and computational budget by introducing an additional full-scale volume.

2.This paper argues that the L2-to-L1 direction provides broader spatial context to the fine-resolution representation, while the L1-to-L2 direction transfers fine-grained evidence to the coarse representation. However, the single-direction results in Table 3 do not support this interpretation.

3.The novelty appears limited, as the framework largely combines standard multi-scale representations with commonly used attention and fusion operations in multi-modal occupancy models. The main distinction lies in the introduction of the L1/L2 spatial partition with different physical resolutions and extents, while the overall methodology follows established design patterns.

Questions for authors
  1. Could the authors justify the fixed-voxel-budget motivation when the proposed framework introduces two full occupancy volumes?

  2. Could the authors further explain why the single-direction GLCI results in Table 3 do not consistently support the claimed roles of L2-to-L1 and L1-to-L2 information transfer?

Reviewer_3723OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UAVOcc to address the resolution–coverage trade-off in large-scale UAV semantic occupancy prediction. The authors also introduce a hierarchical UAV occupancy benchmark based on three existing datasets. Experimental results show that the proposed method achieves consistent improvements in occupancy prediction accuracy over several baseline methods across all three datasets.

Critical strengths

The paper addresses a relevant problem in UAV perception, where both fine-grained local geometry and large-scale spatial coverage are important. The proposed framework is evaluated on three datasets, including real-world UAV observations, and achieves consistent improvements in occupancy prediction accuracy over the compared baselines. The paper is generally well organized, and the overall framework and experimental setup are clearly presented.

Critical weaknesses
  1. Unclear novelty of the proposed hierarchical representation. The proposed L1/L2 representation is motivated by the resolution–coverage trade-off in large-scale occupancy prediction. However, similar problems have already been explored in prior work. For example, OccNeRF (2023) introduced Parameterized Coordinates to address large-scale occupancy estimation through non-uniform spatial parameterization. What is the fundamental novelty of the proposed L1/L2 representation compared with such existing approaches? What specific advantages does this hierarchical representation offer over previously explored non-uniform spatial representations?

  2. Questionable generality of the proposed benchmark. The paper presents the hierarchical UAV occupancy benchmark as one of its main contributions. However, the benchmark appears to be specifically designed around the proposed L1/L2 representation, with two predefined spatial extents and voxel resolutions that exactly match the proposed method. This raises concerns about the generality of the benchmark. Can it fairly and comprehensively evaluate UAV occupancy methods employing different spatial representations, such as uniform voxel grids, continuous implicit fields, or adaptive multi-resolution structures? More fundamentally, if the benchmark is primarily applicable to the proposed L1/L2 hierarchical formulation, it seems more appropriate to regard it as a specially designed experimental setting rather than a general benchmark for UAV occupancy prediction.

  3. Insufficient evidence for the claimed resolution–coverage trade-off. The resolution–coverage trade-off is presented as the central motivation of this work. However, the main paper does not provide sufficient quantitative or qualitative evidence demonstrating that the proposed representation achieves a better trade-off between spatial resolution, physical coverage, and computational efficiency. Although the supplementary material reports computational complexity and runtime comparisons, the proposed method does not consistently outperform the baselines in efficiency. Specifically, Table 11 shows that UAVOcc has an inference latency of 100.84 ms, compared with 16.37 ms for BEVFusion and 41.57 ms for ALOcc. Despite having the lowest reported FLOPs, UAVOcc is substantially slower than these methods in actual inference. Furthermore, the reported FLOPs exclude several operations involved in sampling, warping, and interpolation. The supplementary material also includes an ablation study on latent-grid resolution. However, these experiments only vary the internal feature resolution while keeping the L1/L2 spatial extents fixed. They do not directly evaluate the resolution–coverage trade-off or establish the advantages of the hierarchical representation over alternative spatial representations under comparable computational budgets. Therefore, the experimental evidence does not sufficiently support the paper's central motivation, and relevant controlled efficiency ablations are missing.

  4. Insufficient coverage of related work. The paper contains relatively few references, with limited coverage of recent research, particularly work published in 2025 and 2026. More importantly, the discussion of prior work on large-scale occupancy estimation and non-uniform spatial representations is insufficient.

Questions for authors

What is the fundamental difference between the proposed L1/L2 hierarchical representation and existing approaches such as the Parameterized Coordinates introduced in OccNeRF? What are the specific advantages of the proposed representation?

How general is the proposed benchmark? Can it fairly evaluate UAV occupancy methods that do not adopt the L1/L2 hierarchical representation? What justifies presenting this particular evaluation protocol as a general benchmark rather than a specialized experimental setting?

How does the proposed hierarchical representation improve the resolution–coverage trade-off under a fixed computational or memory budget? Can the authors provide controlled comparisons that isolate the computational advantages of the L1/L2 representation itself?

Reviewer_oUFHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UAVOcc, a hierarchical spatiotemporal semantic occupancy framework for UAV perception. It addresses the resolution–coverage trade-off with a fine-resolution local volume (L1) and a coarse-resolution, larger-range volume (L2). Global–Local Cross-Level Interaction (GLCI) combines metric-aware local correspondence with global contextual retrieval; pose-aligned temporal fusion and Temporal GLCI use historical information across levels. The authors evaluate the framework on UrbanBIS, CUS3D, and UAVScenes, and introduce Package-Physical evaluation for joint spatial coverage. UAVOcc improves Package-Physical mIoU over the strongest listed baselines by 3.41, 5.47, and 5.27 percentage points, respectively.

Critical strengths
  1. The hierarchy represents a physical trade-off relevant to aerial perception.

Unlike a conventional feature pyramid, UAVOcc assigns the two levels different metric resolutions and physical extents. This directly represents the competing needs for local geometric detail and broad spatial coverage, and gives the levels different historical support as the UAV moves.

  1. The evaluation measures both local performance and broader coverage.

The paper establishes a common hierarchical evaluation across three aerial datasets and uses Package-Physical evaluation to avoid double-counting overlapping regions. The L1-only versus L1+L2 comparison also reports gains within the same local physical region, beyond simply extending coverage. The ablations show complementary benefits from spatial interaction, pose-aligned temporal fusion, and coarse-to-fine refinement.

Critical weaknesses
  1. The advantage of the fixed hierarchy over adaptive alternatives is unclear.

Methods such as OctreeOcc already address spatial granularity and representation efficiency. UAVOcc uses distinct physical extents rather than adaptive discretization, but the paper does not establish why this fixed dual-level design is preferable. A comparison with adaptive multi-resolution methods or stronger single-level models under matched voxel and computational budgets would clarify the methodological advantage.

  1. The experiments do not isolate the source of the temporal gains or their cost.

A central motivation for Temporal GLCI is that L2 preserves historical information unavailable in L1. However, the results are reported in aggregate, without separating regions supported by both levels from regions supported only by L2. The gains could therefore reflect additional context or model capacity rather than complementary temporal coverage. The L1+L2 model also adds dense representations and fusion modules, but the paper does not report matched-budget results or measurements such as parameters, FLOPs, latency, and memory, making the accuracy–efficiency trade-off difficult to assess.

Questions for authors
  1. Could you compare UAVOcc with adaptive multi-resolution methods such as OctreeOcc, or stronger single-level models, under matched voxel and computational budgets? Please also report parameters, FLOPs, inference latency, peak GPU memory, and the overhead of Spatial and Temporal GLCI.

  2. Could you report Temporal GLCI’s gains separately for regions supported by both historical levels and regions supported only by L2? This would help determine whether the improvement comes from the proposed complementary temporal coverage.

  3. Could you provide evidence of generalization to unseen geographic environments, clarify the provisional CUS3D baseline records, and evaluate prediction agreement between L1 and L2 in their overlapping region?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.