🌲 SelectAnyTree

A Promptable Instance Segmentation Model for
3D Forest LiDAR Point Clouds

Trung Thanh Nguyen1,2, Daniel Lusk3, Kilian Gerberding3, Janusch Vajna-Jehle3, Tuan-Anh Vu4, Duc Viet Le5, Tu Vo6, Phi Le Nguyen7, Yasutomo Kawanishi8,2,1, Takahiro Komamizu1, Ichiro Ide1, Julian Frey3, Teja Kattenborn3
1Nagoya University 2RIKEN 3University of Freiburg 4UCLA 5University of Twente 6KC Machine Learning Lab 7HUST 8Ritsumeikan University
arXiv Paper Code Demo Data
SelectAnyTree teaser — scene overview with three selected trees, top-down and side view

SelectAnyTree segments any individual tree in a 3D forest LiDAR scan with a user's click. Left: full scene overview with three highlighted trees. Right: zoom panels showing predicted masks. The scene is encoded once; each subsequent click is a cheap query.

Abstract

Instance segmentation of trees in forest LiDAR point clouds is constrained by label scarcity: a single hectare holds millions of points and hundreds of overlapping tree crowns, making manual annotation laborious, while automatic pre-segmentations offer no interactive refinement.

Inspired by the promptable paradigm of foundation segmentation models, we propose SelectAnyTree, a promptable instance segmentation model that delineates any individual tree in a 3D forest point cloud from a few clicks. It couples three lightweight stages: (1) a sparse voxel scene encoder that embeds the forest once into reusable features; (2) a click-to-query prompt encoder that turns each click into a single content query from its 3D position, positive/negative polarity, and the backbone feature of its nearest voxel; and (3) a state-space query decoder that converts this query into one tree mask in linear time, with a mask feedback that conditions each refinement round on the previous mask. Each additional tree costs only a lightweight prompt-encoding and decoding pass, and the full model uses just 19.4M parameters — far fewer than prior promptable 3D models. We further exploit forest-aware geometry by detecting treetops as local maxima of the Canopy Height Model (CHM) and associating one with the user's click as a free initial prompt.

Across seven diverse forest regions and an independent held-out dataset, SelectAnyTree segments a target tree to 79.9 IoU from a single click — 24.7 points above the strongest promptable baseline — and reaches every accuracy target with the fewest clicks.

Demo

Pre-computed from the test set. Each frame shows top-down view (left) and side view (right).

★ Positive click ✕ Negative click ▲ CHM treetop True positive False positive False negative

360° view — select any tree, click by click

Left: the whole plot with every tree segmented. Right: each tree, rotating, as prompts accumulate from 1 to 5 clicks.

SelectAnyTree — 360° rotating view of the scene and each segmented tree as clicks accumulate

Norway (NIBIO)

SelectAnyTree — Norway NIBIO, click-by-click segmentation

Australia (BlueCat)

SelectAnyTree — Australia BlueCat, click-by-click segmentation

Czech Republic (CULS)

SelectAnyTree — Czech Republic CULS, click-by-click segmentation

New Zealand (SCION)

SelectAnyTree — New Zealand SCION, click-by-click segmentation

Point-SAM vs. SelectAnyTree

Same scene, same click budget.

Point-SAM

Point-SAM comparison

SelectAnyTree Ours

SelectAnyTree comparison

Method

SelectAnyTree turns a structure-aware forest backbone into a promptable instance segmentation model through four stages:

  1. Sparse voxel scene encoder (once). The point cloud is voxelized and embedded by a sparse voxel encoder into per-voxel features — cached and reused for all subsequent clicks.
  2. Click-to-query prompt encoder. Each click set is mapped to a single content query from the backbone feature of the click's nearest voxel, a random-Fourier positional encoding of its coordinate, and signed aggregation of positive/negative click polarities.
  3. State-space query decoder. The query is decoded into a voxel-resolution mask by a state-space (Mamba) decoder that captures long-range context in linear time; a lightweight mask feedback conditions each refinement round on the previous mask.
  4. CHM-guided first prompt. Canopy Height Model treetops, detected from scene geometry, provide a "free" first prompt, bridging fully-automatic and interactive segmentation.
SelectAnyTree architecture

SelectAnyTree architecture. Scene encoding is performed once per plot; prompt encoding and decoding are repeated per click set.

Results

79.9 IoU @ 1 click
+24.7 pts above best baseline
#1 fewest clicks at every IoU target
19.4M params — 16× smaller than Point-SAM

Interactive Segmentation

FOR-instanceV2 test set (in-distribution) and LAUTx (cross-dataset generalization, held-out). Bold = best, underline = second best.

Interactive segmentation comparison: IoU@Clicks and NoC@IoU on FOR-instanceV2 and LAUTx

Model Size & Inference Efficiency

GPU wall-clock time (ms) per scene for a k-click session on the FOR-instanceV2 test set.

Model size and inference efficiency comparison

Qualitative Results

Single-click segmentation compared against promptable baselines on the same target tree.

Single-click segmentation qualitative comparison against baselines

BibTeX

@article{nguyen2026selectanytree,
  title   = {SelectAnyTree: A Promptable Instance Segmentation Model for 3D Forest {LiDAR} Point Clouds},
  author  = {Nguyen, Trung Thanh and Lusk, Daniel and Gerberding, Kilian and Vajna-Jehle, Janusch and Vu, Tuan-Anh and Le, Duc Viet and Vo, Tu and Nguyen, Phi Le and Kawanishi, Yasutomo and Komamizu, Takahiro and Ide, Ichiro and Frey, Julian and Kattenborn, Teja},
  journal = {arXiv preprint arXiv:2606.27491},
  year    = {2026}
}