arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

FOSCo‑Net learns to fill forest LiDAR occlusions from a single scan | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
RoboticsEnglishPublished

FOSCo‑Net learns to fill forest LiDAR occlusions from a single scan

November 6, 2026arXiv: 2611.06588v1

FOSCo‑Net is a self‑supervised voxel‑completion framework that reconstructs the geometry hidden behind tree trunks in forest LiDAR data. By accumulating temporally overlapping scans, the system learns a training target that is the set difference between a densely scanned map and a single, sparse input scan, and then predicts where obstacles should be. The network uses a sparse encoder, a dense decoder and σ‑gated skip connections to produce occupancy logits for every voxel.

Single‑scan forest LiDAR image showing a dark, bright‑gap region (blue arrow) behind a tree trunk, depicting an occlusion shadow that is common across multiple scan positions.
Figure 1. Figure 1 illustrates how occlusion shadows behind tree trunks, visible from temporally overlapping scans, provide the automatic supervision needed to define the training target for the self‑supervised voxel‑completion framework.Esfandiyar, Iman, Belter, Dominik · arXiv:2611.06588v1 · Figure 1: · CC BY 4.0

On the DigiForests dataset, FOSCo‑Net was evaluated against dense‑input baselines MIDUNet2D and LMSCNet. With a gated‑skip decoder it reduced the safety‑critical traversability metric SFPR by a factor of 2.1 (0.346 versus 0.741 for MIDUNet2D), while improving other completion metrics such as IoU, F1, recall and precision.

Table of completion metrics for three models: MIDUNet2D, LMSCNet, and FOSCo‑Net. The row for FOSCo‑Net reports the lowest SFPR value and higher IoU, F1, recall, and precision compared to the other models.
ModelGridSFPR↓IoU↑F1↑Rec↑Prec↑
MIDUNet2D‡ [11]Cyl.0.7410.1570.2680.7100.159
LMSCNet‡ [9]Cart.−0.1190.2100.6920.141
Ours (FOSCo-Net)Cyl.0.3460.1720.2900.6960.187
Cart.−0.1810.3030.6830.193
Table 2. Table 2 compares FOSCo‑Net (Gated Skip) with dense‑input baselines MIDUNet2D and LMSCNet, showing a 2.1× reduction in the safety‑critical traversability metric (SFPR 0.346 versus 0.741 for MIDUNet2D).Esfandiyar, Iman, Belter, Dominik · arXiv:2611.06588v1 · Table 2: · CC BY 4.0

The framework’s performance is limited to a single sequence of 50 ego‑centric sample pairs from the DigiForests dataset; its generalizability to other forest types or larger‑scale tests remains unverified.

Source paper

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

Robotics
View on arXiv