acceptodds
Under review as a conference paper at ICLR 2027

EndoGeoNet: Geometry-Arbitrated Thin-Branch Segmentation for Forest Navigation

Abstract

Thin branches are difficult to segment in forest scenes because they occupy few pixels and often share appearance with surrounding vegetation. We introduce EndoGeoNet, a monocular segmentation framework that augments a strong RGB backbone with geometry estimated from the same input image. A frozen monocular depth estimator provides a relative-depth cue, which is encoded by a lightweight multiscale geometry branch and fused at four decoder stages through Geo-Visual Arbitration Modules (GVAMs). Each GVAM learns stage-specific appearance–geometry weights, allowing the model to suppress unreliable depth while retaining useful geometric corrections. We also introduce the Forestry Sparse-Branch Dataset (FSB), a dedicated benchmark of 60 manually annotated full-HD images acquired across different locations and capture batches, together with topology- and component-aware evaluation. On FSB, EndoGeoNet improves the RGB baseline from 0.587 to 0.629 Dice and from 0.537 to 0.580 clDice, while maintaining near-identical performance on DIS-5K (0.899 versus 0.900 Dice). Replacing aligned depth with zero or shuffled depth reduces Dice by 2.19 and 2.63 points, respectively, showing that the model relies on spatially corresponding geometry. Including online depth estimation, the full pipeline runs at 7.74 FPS in FP32 and 16.74 FPS with mixed precision on an RTX 4090.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Endogenous geometry, thin-branch segmentation, monocular depth estimation, geometry-aware segmentation, dichotomous image segmentation, forest navigation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_g9QKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EndoGeoNet augments a frozen RGB segmentation model with monocular pseudo-depth through a lightweight geometry branch and stage-wise fusion gates. It introduces a 60-image forest-branch benchmark and reports improved thin-structure segmentation after training and model selection on DIS-5K.

Critical strengths

The method addresses a concrete thin-structure segmentation problem, and the evaluation includes centerline and component-level metrics relevant to branch recovery. FSB is excluded from training and model selection, making the reported transfer results useful.

Table 4 controls the backbone, depth cues, training data, and optimization. GVAM outperforms the reported concatenation alternatives while using fewer trainable parameters, supporting its practical efficiency.

Critical weaknesses

The FSB results require reconciliation. Table 2 reports Dice = 0.629, whereas the aligned-depth condition in Table 3 reports 0.649 and GVAM in Table 4 reports 0.6492. No corresponding checkpoint or evaluation-condition distinction is stated. This difference is comparable to the 0.0219 depth-ablation gain in Section 5.2, making clarification necessary for interpreting the reported improvements.

The evidence supports geometry utilization more directly than reliability-aware arbitration. The abstract claims suppression of unreliable depth. Section 5.2 establishes dependence on spatially aligned geometry, and Table 4 supports GVAM over concatenation, but neither directly isolates image-adaptive weighting or demonstrates suppression of unreliable depth. This leaves the distinctive role of adaptive arbitration insufficiently established.

Questions for authors
  1. What accounts for the FSB Dice scores of 0.629, 0.649, and 0.6492 in Tables 2–4? Do these results use different checkpoints, seeds, or evaluation conditions?
  2. Does GVAM outperform fixed fusion weights or image-independent learned weights when the geometry branch and training protocol are otherwise unchanged?
  3. Beyond the zero-depth and shuffled-depth results, what evidence shows that GVAM reduces reliance on unreliable geometry? Under controlled depth corruption, how do the fusion weights and segmentation performance change?

I would raise my rating if the authors reconcile the reported results and substantiate the benefit of adaptive geometry fusion, and lower it if corrected results or matched controls materially weaken the central contribution.

Reviewer_QTHbOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies monocular thin-branch segmentation in forest scenes. EndoGeoNet augments a frozen SAM2-UNeXT segmentation model with relative depth from frozen Depth Anything V2 Small, a four-layer geometry encoder, and four Geo-Visual Arbitration Modules (GVAMs). Only the geometry encoder and fusion modules are trained, using DIS-TR for training and DIS-VD for model selection. Evaluation covers four DIS test splits and the authors' 60-image Forestry Sparse-Branch Dataset (FSB). The paper reports largely unchanged general DIS performance and improvements over the RGB baseline in FSB Dice, clDice, and component miss rate.

Critical strengths
  1. The task is practically relevant and the evaluation objective is clear. Foreground segmentation models can miss thin forest branches. The paper distinguishes training on DIS from evaluation on FSB without target-domain fine-tuning, and describes annotation rules for visible woody foreground and an independent re-annotation check on five images (PDF p.4, Sec.4; p.11, Table7). These provide a useful starting point for studying this application.

  2. The additional trainable component is small, and full inference costs are reported. The geometry encoder and fusion modules reportedly contain 352,840 trainable parameters, while the original appearance encoder and decoder remain frozen. Online timing includes depth estimation and segmentation and is explicitly scoped to an RTX 4090 (PDF p.3–4, Sec.3; p.8, Table5).

Critical weaknesses
  1. Algorithmic novelty is limited, and a substantive advance beyond existing pseudo-depth fusion and attention weighting is not established. The core additions are a four-layer standard convolutional geometry encoder and GVAM's sequence of 1×1 alignment, concatenation, global average pooling, a two-layer MLP, two scalar softmax weights, and weighted summation (PDF p.3–4, Eqs.(2)–(6)). PDFNet already generates Depth Anything V2 pseudo-depth from RGB and performs multistage fusion for the same DIS task. Selective Kernel Networks already use global pooling, compact projection, and softmax branch weighting. DPANet already studies gated fusion that reduces the adverse effect of unreliable depth on segmentation.

  2. Incomplete baseline configurations limit reproducibility of the cross-domain comparison. Table2 gives SAM2-UNeXT a Dice score of 0.587 and the other six baselines 0.074–0.306 (PDF p.6). The manuscript specifies no target-domain fine-tuning, no test-time augmentation, and a threshold of 0.5, but does not identify each method's checkpoint, training source, input processing, or prediction processing; PDFNet's depth generator is also unspecified. These configurations are needed to reproduce the rankings in Tables1/2 and explain the cross-domain differences. For example, the official SAM2-UNeXT script applies per-image min–max normalization after sigmoid, which affects results at a threshold of 0.5; the manuscript does not identify the processing actually used. Please provide per-baseline configurations and a common metric computation procedure, including PDFNet's required depth input.

  3. FSB is small and lacks representativeness and stability analyses supporting its benchmark value and model rankings. FSB contains only 60 images, is presented as a separate contribution, and provides the main evidence of performance improvement (PDF p.2, contribution list; p.4, Sec.4; p.6, Table2).

Questions for authors

Please see Weakness.

Reviewer_Ws6GOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a monocular segmentation framework that targets forest branches, through multi scale depth estimation and four stage decoder fusion. It also introduces a 60 image segmentation benchmark with topology aware metrics.

Critical strengths

The paper introduces a practical benchmark FSB, with a clear annotation protocol. the evaluation metrics used like Skeleton F1(missing in table 1&2), miss rate and cIDice are appropriate.

Critical weaknesses
  • Baselines compared to in table 2, do not get equivalent training budget, model parameters or pretaining as EndoGeoNet. also the backbone used SAM2UNeXT is a strong backbone for the task. A good baseline would replicate some of the baseline methods with the backbone used in EndoGeoNet
  • EndoGeonet numbers reported in table 2 and 3 have a mismatch. If the numbers in table 2 are true, that would mean the depth input doesn't influence the model.
Questions for authors

address the weaknesses above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.