FOSCo-Net: Ego-centric Voxel Completion for Forest LiDAR Occlusions

Iman Esfandiyar; Dominik Belter

cs.RONov 4, 2026

FOSCo-Net: Ego-centric Voxel Completion for Forest LiDAR Occlusions

Authors: Iman Esfandiyar, Dominik Belter

Organizations: Poznań University of Technology, Institute of Robotics and Machine Intelligence, Poznań, Poland · Łukasiewicz Research Network – Poznań Institute of Technology, Center of Digital Transformation, Department of Artificial Intelligence, Poznań, Poland

Abstract

Ground robots operating in temperate forests suffer from systematic LiDAR occlusions caused by tree trunks, creating persistent blind regions that degrade traversability estimation. In this paper, we propose FOSCo-Net, a self-supervised voxel completion framework that learns to reconstruct occluded geometry from a single scan using temporally accumulated observations as automatic supervision, and we evaluate cylindrical and Cartesian ego-centric grids with multiple decoder designs. FOSCo-Net consistently improves reconstruction quality and traversal safety, with a gated skip decoder achieving the most reliable performance and significantly reducing false obstacle predictions compared to baselines.

Figures & tables

Figure 1: Single LiDAR scan in a temperate forest. The bright gap (blue arrow) behind a tree trunk is an occlusion shadow.

Figure 2: Self-supervised training and inference pipeline. Top: LiDAR frames from a temporal window

[T{-}\Delta,\,T{+}\Delta]

are registered into the frame at

T

to form the dense supervision target. Bottchaom: at inference only the single scan at

T

is available.

Dimension	Cylindrical	Cartesian	Range	Resolution
Radial $r$ / $x$	40	120	2–6 m ∗ / $\pm$ 6 m	0.10 m
Angular $\theta$ / $y$	360	120	0°–360° ∗ / $\pm$ 6 m	1° ∗ / 0.10 m
Vertical $z$	250	250	$-$ 5–20 m	0.10 m
Total cells			3,600,000

Table 1: Grid parameters for both representations ( ∗ cylindrical only). Both grids share the same vertical extent and 0.1 m resolution, yielding 3.6M total cells.

Panel of Figure 3 — Figure 3: Comparison between Ego-centric grid pairs: cylindrical input (cyan = ground, yellow = stem, brown and green points are not used by both representation) (a) and target output (b),Cartesian input (c) and target output (d). The model predicts voxels in the target absent from the input.

Panel of Figure 4 — Figure 4: Proposed decoders based on the same sparse encoder (yellow/banded blocks): implicit, with a coordinate-conditioned $1{\times}1{\times}1$ MLP head (teal) (a), full Sparse, using SparseConvTranspose3d upsamplers (orange) (b), dense Decoder, with trilinear DenseUpBlocks after an early .dense() (c), Gated Skip, extending the Dense Decoder with $\sigma$ -gated skip connections (purple arcs) from each encoder level (d).

Model	Grid	SFPR $\downarrow$	IoU $\uparrow$	F ${}_{1}\uparrow$	Rec $\uparrow$	Prec $\uparrow$
MIDUNet2D ‡ [ 11 ]	Cyl.	0.741	0.157	0.268	0.710	0.159
LMSCNet ‡ [ 9 ]	Cart.	$-$	0.119	0.210	0.692	0.141
Ours (FOSCo-Net)	Cyl.	0.346	0.172	0.290	0.696	0.187
Ours (FOSCo-Net)	Cart.	$-$	0.181	0.303	0.683	0.193

Table 2: FOSCo-Net (Gated Skip) vs. dense baselines (worst to best,

\tau{=}0.5

). ‡ Dense-input, no sparse encoder. Column defs and colour scale as Table 3 . SFPR, the safety-critical traversability metric, is reduced by a factor of 2.1 compared with MIDUNet2D.

		Trav.	All-voxel		Stem	Pred. qual.
Architecture	Grid	SFPR $\downarrow$	IoU $\uparrow$	F ${}_{1}\uparrow$	Rec $\uparrow$	Prec $\uparrow$
Implicit †	Cyl.	0.344	0.161	0.276	0.508	0.164
Implicit †	Cart.	$-$	0.149	0.257	0.453	0.182
Full Sparse	Cyl.	0.336	0.165	0.282	0.405	0.226
Full Sparse	Cart.	$-$	0.169	0.288	0.397	0.222
Dense Dec.	Cyl.	0.359	0.170	0.287	0.668	0.184

Table 3: FOSCo-Net ablation on 10 held-out testing samples (

\tau{=}0.5

). All eight variants (four decoders

\times

two grids); ordered worst to best by all-voxel accuracy. Colour scale: best worst per column. Rec = stem recall; Prec = global precision; SFPR = cylindrical-only shadow FPR (lower is better). Bold : best per column (SFPR: recall-competitive models,

\mathrm{Rec}_{\text{stm}}{>}0.65

). † Implicit: batch size 1.

Panel of Figure 5 — Figure 5: FOSCo-Net (Gated Skip) predictions overlaid on the input scan (a,c) (magenta = predicted ground, orange = predicted stem): cylindrical grid (a,b) and Cartesian grid (c,d).

Panel of Figure 6 — Figure 6: Dense baseline predictions, same scene as Fig. 5 (magenta = predicted ground, orange = predicted stem):Make-it-Dense (MIDUNet2D) [ 11 ] (a,b) and LMSCNet [ 9 ] (c,d). For both methods trunk locations are recovered but predictions smear across neighbouring voxels, consistent with the low precision ( $0.141$ ) in Tab. 2 .

Point Cloud Completion

Point cloud completion reconstructs missing 3D geometry from sparse, partial, noisy, or occluded point cloud observations.

See all 32 papers

Predicting Void Points in Self-Occluding Forest Volumes
Nov 3, 2026Benjamin Bachmayr, Mohamed Youssef, Oliver BimberForest Remote Sensing 3D Occupancy Prediction
Institute of Computer Graphics, Johannes Kepler University Linz, Austria
ZVeC: A Zero-Shot Framework for Instance-Level Vehicle Extraction and Generative Point Cloud Completion
Sep 21, 2026Daisy Li, Kyle Gao, Quanyun Wu +3Point Cloud Completion 3D Scene Completion
University of Waterloo · Karlsruhe Institute of Technology
GhostPoint: Self-Supervised Representation Learning by Hallucinating Occluded LiDAR Structure
Aug 14, 2026Mohamed Abdelsamad, Bin Yang, Michael Ulrich +4Self-Supervised Point Cloud Learning Lidar-Based 3D Object Detection
Bosch Center for AI · University of Freiburg · University of Luebeck

Nov 3, 2026J/K move · Enter open · S save

Predicting Void Points in Self-Occluding Forest Volumes

Benjamin Bachmayr, Mohamed Youssef, Oliver Bimber

Institute of Computer Graphics, Johannes Kepler University Linz, Austria

Forest Remote Sensing 3D Occupancy Prediction 3D Reconstruction Remote Sensing

Open paper AskPDF arXiv

Synthetic Aperture Imaging (SAI) enables drones to observe forest canopies by warping and averaging multiple displaced views at a common depth. However, occlusion is suppressed rather than eliminated, causing poorly observed regions to appear occupied and leading to systematic biomass overestimation. The Deep Forest project cannot distinguish reconstruction artifacts from real vegetation because its training data relies on measured reflectance points, excluding empty space and preventing the network from learning explicit occupancy boundaries. This thesis addresses the problem by modifying training labels rather than the imaging process. Each simulated scene is rendered twice: as a 9x9 grid of top-down captures for focal-stack generation and as 550 depth-clipped slabs derived from scene geometry to provide occupancy labels, explicitly representing void space as a negative class. A Gazebo pipeline generates procedural forest stands containing 80-200 trees across three species. A pretrained Pix2Pix model translates rendered RGB images into a synthetic near-infrared band, and Light Field Rendering produces aligned focal stacks of 550 planes across three channels. Four architectures were evaluated on a held-out scene. The plain 2D U-Net achieved an IoU of 0.6290, outperforming the volumetric 3D U-Net (0.5707). Established reconstruction methods performed substantially worse: COLMAP achieved 0.3446 and DepthAnything3 0.2959, both after their respective optimal dilation sweeps. An RGB-input variant of the volumetric architecture achieved the highest IoU of 0.6514 without dilation or per-scene calibration. These results highlight the limitations of conventional reconstruction methods, which recover visible surfaces but struggle to infer occupancy within closed canopies.

Abstract

Point Cloud Completion

Explore similar work

Predicting Void Points in Self-Occluding Forest Volumes

ZVeC: A Zero-Shot Framework for Instance-Level Vehicle Extraction and Generative Point Cloud Completion

GhostPoint: Self-Supervised Representation Learning by Hallucinating Occluded LiDAR Structure