Neurocomputing · 2026

NeRF-UAVeL

Unified attention-driven volumetric learning for robust NeRF-based 3D object detection

Hana L. Goshu, Tadesse G. Wakjira, Meklit M. Atlaw, Kin-Chung Chan, Songjiang Lai, Kin-Man Lam

Paper Code Project page Data

NeRF-UAVeL

Ground truth

Ground truthDrag to compareNeRF-UAVeL

AP50 over NeRF-RPN on 3D-FRONT: +6.7
R50 over NeRF-RPN on 3D-FRONT: +7.3
AP50 over NeRF-RPN on ScanNet: +6.9

Abstract

Three-dimensional (3D) object detection based on Neural Radiance Fields (NeRF) has emerged as a promising direction for reconstructing complex environments from posed RGB images. However, existing NeRF-based detectors often suffer from coarse feature encoding and limited attention to multi-scale volumetric structures, leading to inaccurate localization and poor generalization in real-world scenarios. To address these challenges, we propose NeRF-UAVeL, a unified attention-driven volumetric learning framework for 3D object detection that integrates four novel modules into a NeRF-derived 3D volumetric backbone, namely, Multi-dimensional Volumetric Attention Pooling (MVAP), Tri-Scale Asymmetric Convolutional Aggregation (TACA), Dual-Domain Attention Fusion (DDAF), and Volumetric Cross-Window Attention Fusion (V-CWAF). MVAP enhances spatial selectivity through adaptive attention-based pooling, TACA captures multi-scale volumetric features through asymmetric convolutional branches, DDAF applies channel and multi-scale spatial recalibration for refined feature emphasis, and V-CWAF applies windowed self-attention with dual-stage channel recalibration to boost high-level semantic encoding. Extensive experiments on the 3D-FRONT and ScanNet datasets demonstrate that NeRF-UAVeL outperforms both point cloud-based and multi-view-based methods. Specifically, it improves AP50 by 6.7 and R50 by 7.3 over the baseline on 3D-FRONT, and achieves a 6.9 improvement in AP50 and 2.9 in R50 on the ScanNet dataset. These results confirm the effectiveness of our attention-calibrated, multi-scale volumetric architecture in producing precise and robust 3D bounding box predictions across both synthetic and real-world scenes.

Method

NeRF-UAVeL augments a NeRF-derived volumetric backbone with four attention-driven modules at different stages of the feature hierarchy.

MVAP (Multi-dimensional Volumetric Attention Pooling) enhances spatial selectivity through adaptive attention-based pooling across volumetric dimensions.
TACA (Tri-Scale Asymmetric Convolutional Aggregation) captures multi-scale volumetric features through asymmetric convolutional branches at three complementary scales.
DDAF (Dual-Domain Attention Fusion) applies channel and multi-scale spatial recalibration for refined feature emphasis.
V-CWAF (Volumetric Cross-Window Attention Fusion) applies windowed self-attention with dual-stage channel recalibration to strengthen high-level semantic encoding.

NeRF-UAVeL architecture — From multi-view images to 3D detection: the four attention modules sit at different stages of the volumetric feature pyramid.

Results

The first three methods are point cloud-based; the remaining methods detect objects from multi-view images.

3D object detection on 3D-FRONT and ScanNet (recall and average precision at IoU 0.25 and 0.5).

Method	3D-FRONT R25	R50	AP25	AP50	ScanNet R25	R50	AP25	AP50
VoteNet	81.5	61.6	73.0	49.6	78.5	34.2	66.8	18.2
GroupFree	84.9	63.7	72.1	45.1	75.2	37.6	60.1	20.4
FCAF3D	89.1	56.9	73.1	35.2	90.2	42.4	63.7	18.5
ImVoxNet	88.3	71.5	86.1	66.4	51.7	20.2	37.3	9.8
NeRF-MAE*	97.2	74.5	85.3	63.0	92.0	39.5	57.1	17.0
NeRF-RPN	96.3	69.9	85.2	59.9	89.2	42.9	55.5	18.4
NeRF-UAVeL	98.5	77.2	87.3	66.6	92.6	45.8	60.1	25.3

* NeRF-MAE is pretrained on 3D-FRONT and two additional datasets, Hypersim and HM3D.

Qualitative detection results — Predicted boxes on 3D-FRONT scenes: input image, NeRF-RPN, NeRF-UAVeL and ground truth.

BibTeX

@article{goshu2026nerfuavel,
  title   = {NeRF-UAVeL: Unified attention-driven volumetric learning for robust NeRF-based 3D object detection},
  author  = {Goshu, Hana L. and Wakjira, Tadesse G. and Atlaw, Meklit M. and Chan, Kin-Chung and Lai, Songjiang and Lam, Kin-Man},
  journal = {Neurocomputing},
  volume  = {702},
  pages   = {134610},
  year    = {2026},
  doi     = {10.1016/j.neucom.2026.134610}
}