Beyond Appearance: Where Forest Visual Place Recognition Actually Fails

Blake Matheson, Ben McShanag, Isabella Moffat, Michael Milford, Peyman Moghadam, Dana Kulić, Somayeh Hussaini, Kavindie KatuwandeniyaarXiv 2026 · arXiv

TL;DR

This paper rigorously investigates whether non-appearance cues—temporal, linguistic, structural, and geometric—can solve the "boss level" of Visual Place Recognition (VPR): repetitive, ever-changing forest environments. The authors find that while temporal sequence matching significantly boosts high-precision localization, other modalities like VLM captions and scene graphs currently struggle to provide discriminative information beyond what a strong appearance-only baseline already captures.

What Problem Does This Paper Address?

Visual Place Recognition (VPR) is the task of a robot recognizing a previously visited location using only camera data. In urban environments, we have unique buildings and signs. In forests, we have "perceptual aliasing": every trail looks like every other trail, and the same spot looks completely different after a season change or when approached from the opposite direction.

Current state-of-the-art VPR models rely heavily on global appearance descriptors. While powerful, these models often fail when lighting shifts or viewpoints change drastically. This paper asks a fundamental question: Can we go "beyond appearance" by using context? Specifically, can we use the order of images (temporal), what’s in the scene (linguistic/structural), or the physical layout of trees (geometric) to fix the failures of appearance-based models?

Methods

The researchers established a strong appearance-only baseline using MegaLoc and then branched out into four "diagnostic axes" to see what added value they provided:

Figure 1: The four diagnostic axes, each defined by the space in which similarity is measured. The temporal axis acts on the appearance pipeline itself, aggregating either the descriptors or the similarity matrix; the linguistic, structural, and geometric axes branch from the frames and produce their own similarity matrices. A gate on the similarity matrix S selects queries whose shortlist is reordered by a reranker ρ on images. Each output is compared against the baseline S on identical traversals.
  1. Temporal Context: They compared two training-free sequence matching strategies. Score-space matching (like SeqSLAM) looks for consistent high-similarity paths in the similarity matrix. Descriptor-space matching (like SeqNet) averages descriptors over a window before matching.
  2. Linguistic Context: They used Vision-Language Models (VLMs) like Qwen2.5-VL and GPT-5.6 Sol to generate detailed captions of forest scenes. They then used text-retrieval metrics to see if these descriptions were unique enough to identify a place.
  3. Structural Context: They built Scene Graphs by detecting objects (trunks, canopy, trails) using Grounding-DINO and SAM2, then encoding their spatial relationships and appearances with DINOv2.
  4. Geometric Context: They created Stem Graphs—top-down maps of tree trunk positions—by combining SAM3 segmentation with monocular depth estimation (Depth Anything 3). These were matched using Iterative Closest Point (ICP) or Bird's-Eye-View (BEV) descriptors.
The process of constructing stem graphs from masks and depth maps.

Datasets

The study utilized the WildCross benchmark, which features challenging forest traversals in the Venman and Karawatha environments. The dataset includes:

  • Same-direction revisits: Standard loop closure.
  • Extended routes: Where the robot takes a slightly different path.
  • Reverse traversals: Approaching the same place from the opposite direction (the hardest case).
  • Long-term revisits: Returning to a site ~14 months later.

Key Insights and Ideas

  • Temporal Context is the Only Consistent Winner: Sequence matching effectively filters out "isolated" false positives—cases where a single frame looks like a different place but the surrounding frames do not.
  • The Precision-Recall Trade-off in Sequences: Score-space matching is superior for high-precision tasks (like loop closure in SLAM). While descriptor-space matching reaches similar Recall@1, it "blurs" distinct places together, causing a massive drop in recall at 100% precision.
  • Forests Break Language: VLM captions proved too generic. A description like "a dirt path surrounded by eucalyptus trees" applies to hundreds of meters of trail, leading to poor place-specific discrimination.
  • Geometry is Fragile: While tree trunks are persistent landmarks, monocular depth estimation is currently too noisy to create reliable metric maps for matching. Even using LiDAR-derived depth only marginally improved results.
  • Reverse Traversals Remain Unsolved: When facing the opposite way, the visual overlap is often so low that even temporal aggregation can't find enough common ground to recover a match.

Key Results

  • Baseline Performance: MegaLoc achieved a solid ~79% Recall@1 on same-direction revisits.
  • Temporal Boost: On a 14-month revisit (V-04), score-space sequence matching raised Recall@1 from 87.5% to 94.1%. More importantly, it boosted Recall at 100% Precision (R@100P) from 50.9% to 71.6%.
  • Context Failures: Linguistic, structural, and geometric streams generally performed far below the appearance baseline. For example, the best caption-based retrieval reached only ~22% Recall@1, compared to the baseline's ~87%.
Figure 3: Precision-recall and R@K plots for MegaLoc on the V-01/V-04 dataset (values as fractions in [0, 1]).

Limitations and Future Work

The authors acknowledge several hurdles:

  • Zero-Shot Constraints: The VLM and scene graph components were used "out of the box." Fine-tuning these models specifically on forest data might unlock more discriminative power.
  • Depth Accuracy: The failure of stem graphs was largely due to the "hallucinations" and errors in monocular depth. Future work could integrate more robust stereo or LiDAR-based geometric features.
  • Viewpoint Invariance: To solve reverse traversals, we need representations that are fundamentally more invariant to 180-degree rotations than current global descriptors.

In summary, if you're building a robot for the woods today, stick to appearance-based descriptors and layer on score-space sequence matching. The "fancier" modalities like language and scene graphs aren't quite ready for the trail yet.

Summary updated

Abstract

Visual place recognition (VPR) in forest environments remains challenging because repetitive vegetation, and changes in illumination, season, and viewpoint can make distinct places appear similar and revisited places appear substantially different. In this paper, we investigate whether temporal, linguistic, structural, and geometric context can improve on an appearance-only descriptor, and what limits each representation when it does not. Using MegaLoc as the appearance-based baseline on the WildCross benchmark, we evaluate score-space and descriptor-space sequence matching, vision-language captions and reranking, scene-graph representations, and depth-derived stem graphs across same-direction revisits, extended routes, and reverse traversals. On a same-direction revisit, both sequence matching strategies reach similar Recall@1 (94.1% vs. 91.5%), but recall at 100% precision rises from 50.9% (single-frame MegaLoc) to 71.6% in score space and falls to 36.5% in descriptor space. The remaining context representations provide no consistent improvement over the appearance-based baseline. Same-place captions show limited place-specific discrimination, reranking is constrained by the initial candidate shortlist, and stem-graph matching remains weak even with LiDAR depth.
BibTeX
@misc{matheson2026beyond,
  author = {Blake Matheson and Ben McShanag and Isabella Moffat and Michael Milford and Peyman Moghadam and Dana Kulić and Somayeh Hussaini and Kavindie Katuwandeniya},
  title = {{Beyond Appearance: Where Forest Visual Place Recognition Actually Fails}},
  year = {2026},
  eprint = {2611.21996},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21996},
}
Figure