Pith. sign in

REVIEW 3 major objections 7 minor 12 references

Parameter-Efficient Fine-Tuning of Vision Foundation Model for Forest Floor Segmentation from UAV Imagery

T0 review · 3 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Adapter-based parameter-efficient fine-tuning gives the best forest-floor segmentation from UAV imagery, with LoRA nearly matching it at half the trainable parameters.

desk verdict A straightforward, honest comparison of PEFT adapters vs. LoRA for SAM on a new UAV forest-floor dataset—useful as a workshop study, but the 0.006 mIoU gap between the top two methods needs much stronger evidence before anyone should lean on it. read the letter →

arxiv 2505.08932 v1 pith:PDLY57R5 submitted 2025-05-13 cs.RO cs.CV

Mohammad Wasil , Ahmad Drak , Brennan Penfold , Ludovico Scarton , Maximilian Johenneken , Alexander Asteroth , Sebastian Houben This is my paper · ORCID
classification cs.ROcs.CV
keywords forestfloorsegmentationUAVimagerySegmentAnythingModelparameter-efficientfine-tuningLoRAadaptertuningcoarsewoodydebrisorthomosaic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether a vision foundation model, SAM, can be adapted to segment forest-floor objects in UAV orthomosaics—tree stumps, coarse woody debris, and vegetation—without retraining the entire model and without manual prompts. It reports that injecting small trainable modules into the frozen image encoder while fine-tuning only the mask decoder is enough: the serial-adapter variant reaches mIoU 0.416, LoRA reaches 0.410 with roughly half the trainable parameters, and both clearly beat fine-tuning the mask decoder alone (0.321) and zero-shot SAM (0.151). The value would be a practical route to automatic obstacle detection for UAV-assisted reforestation, where model size and compute are constrained.

What carries the argument

The mechanism is a set of PEFT modules placed in the frozen image encoder of SAM, combined with a modified mask decoder. Adapters are bottleneck MLPs (down-projection, nonlinearity, up-projection) inserted after attention or beside the MLP, with a scale factor on the residual. LoRA adds low-rank matrices B and A to the query and value weight matrices, W_q and W_v. The paper keeps the image encoder frozen, trains only the PEFT modules, and replaces SAM's prompt-dependent decoding so the decoder outputs N class masks for N categories, removing the prompt encoder entirely.

What would settle it

Re-annotate the 36 test windows with a stricter, consensus-based definition of coarse woody debris and with stumps separated from upside-down logs, then retrain all four variants under the same protocol; if the serial adapter no longer leads LoRA, or if the order reshuffles across random seeds, the reported ranking is an artifact of annotation noise.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that PEFT can take SAM from a prompt-driven generic segmenter to an automatic class-specific segmenter for a difficult natural scene. With SAM's image encoder frozen, adding two serial bottleneck adapters per transformer block yields mIoU 0.416; LoRA added to the query and value projections yields 0.410 from 4.99M trainable parameters; a single post-MLP adapter yields 0.359; training only the mask decoder yields 0.321; and zero-shot SAM with point prompts yields 0.151. The paper concludes that adapter- and LoRA-based tuning outperform direct decoder fine-tuning, that where the modules sit matters more than parameter count, and that LoRA is the lightweight option for resource-constrained platforms.

Load-bearing premise

The results depend on the forest-floor annotations being accurate enough that the measured mIoU differences (notably 0.416 versus 0.410) reflect method quality rather than label noise, a premise the paper itself weakens by reporting many missed or mislabeled coarse-woody-debris instances and only 36 test windows.

Editorial extensions

If this is right

  • Forest-floor obstacle maps can be produced automatically from UAV orthomosaics without manual prompting.
  • LoRA, with 4.99M trainable parameters, is a viable lightweight choice for on-board deployment on resource-constrained UAV platforms.
  • PEFT avoids full fine-tuning of SAM, cutting memory and compute requirements.
  • Performance is uneven across classes, with background dominating and coarse woody debris the lowest, pointing to annotation quality as the next bottleneck.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • In our reading, the 0.006 mIoU gap between the serial adapter and LoRA is within the noise floor of a 36-window test set with known label inconsistency, so the ranking should not be treated as stable until labels are cleaned.
  • The paper's finding that module placement matters more than parameter count suggests a testable extension: sweeping adapter positions across attention and MLP blocks to find a sparser configuration that still matches the serial adapter.
  • Extending the same protocol to the other unannotated fields of the surveyed forest area would show whether the ranking transfers across sites and flight conditions.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The manuscript adapts the Segment Anything Model (SAM) to forest floor segmentation from UAV orthomosaics using parameter-efficient fine-tuning (PEFT). The image encoder is frozen; AdapterH, AdapterL, LoRA modules, or no-PEFT (decoder-only) are trained together with a modified mask decoder that outputs class masks without prompts. Evaluation is on field-D of the Garrulus dataset (0.3 ha, ~10 cm/pixel), with 1,000 training windows and 36 fixed-grid test windows, three seeds. Table I reports AdapterH mIoU 0.416, LoRA 0.410, AdapterL 0.359, decoder-only 0.321, and zero-shot SAM 0.151. The paper concludes that adapter-based PEFT achieves the highest mIoU, with LoRA as a lightweight alternative for UAV platforms.

Significance. If the ranking is robust, the paper gives a practical recipe for adapting a vision foundation model to a small-domain remote sensing task with limited trainable parameters, and the code release plus the honest discussion of label noise are commendable. The main contribution is empirical and domain-specific; it does not claim new PEFT theory. The result is potentially useful for reforestation and seed-dispersal pipelines, but its strength is bounded by the single 0.3 ha field, the 36-window test set, the acknowledged annotation ambiguity, and the lack of a full-fine-tuning baseline.

major comments (3)
  1. [§IV-A, Table I] The headline comparison between AdapterH (mIoU 0.416) and LoRA (0.410) rests on a 0.006 difference evaluated on 36 fixed-grid windows from a single orthomosaic. The reported ±0.001 standard deviations are across three training seeds and do not capture test-window sampling variability or label noise, which Section IV-C explicitly documents as severe for CWD (incorrectly labeled, missed instances, subjective boundaries). With only 36 test windows, a bootstrap or permutation test over windows is needed to show that the AdapterH-vs-LoRA ordering is not a sampling artifact, and ideally the evaluation should be repeated after correcting or excluding the most ambiguous CWD labels. Without this, the statement that AdapterH achieves the highest mIoU is not supported to the precision implied by Table I.
  2. [§IV-C, Fig. 10] mIoU is reported as a single macro-average, but Fig. 10 shows that the BACKGROUND class dominates the per-class results, and MISC and STUMP have significantly lower mIoU. Because the target application is obstacle detection (CWD, STUMP, vegetation), a method could win on the background-dominated macro mIoU while being worse on exactly the classes that matter. The paper should report per-class mIoU in a table for all four classes and, if the macro average is kept, also give a foreground-only mIoU. This would clarify whether the AdapterH/LoRA advantage is concentrated in the classes relevant to the seeding application.
  3. [§IV-C, zero-shot protocol] The zero-shot SAM baseline is not measured on the same task as the supervised methods. SAM's automatic mask generator produces class-agnostic proposals; the authors then decide a predicted mask is a true positive if its IoU with a ground-truth mask exceeds 50%, and this is aggregated into an mIoU. That is a detection and proposal matching heuristic, not semantic segmentation, and the 50% threshold is an additional design choice. The 0.151 figure should be labeled as a proxy, or replaced by a class-agnostic metric such as average recall of proposals or mask IoU without class assignment, so that the comparison with the PEFT mIoU values is not misleading.
minor comments (7)
  1. [§IV-B] The learning-rate description 'cosine annealing schedule that decreased the rate by a factor of 0.1 at each step' is inconsistent with cosine annealing; please describe the schedule and the final learning rate precisely.
  2. [§IV-B] The loss weighting between weighted cross-entropy and Dice loss is not given; please specify the weights used.
  3. [Table I] The caption does not state that the ± values are standard deviations over three training seeds; please state this and, ideally, provide per-window mIoU variability as well.
  4. [§IV-C] The dense-embedding ablation reports improvements of '0.02% and 0.04%'; these appear to be absolute mIoU percentage points rather than relative percentages, and the text should state which one is meant.
  5. [§IV-C, references] The zero-shot experiment is described as 'following the approach by Hu et al. [7]', but [7] is the LoRA paper and is not the source for SAM's automatic mask generator; please correct the citation.
  6. [§IV-A] The dataset is not yet released, and the test set consists of 36 windows; even a small release of the test annotations or a data-generation script would materially improve reproducibility.
  7. [Title and §IV-C] There are small typos: 'UA V' appears with a spurious space in the title and abstract, and 'usign' appears in Section IV-C in the ablation description.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the paper is an empirical benchmark whose scores are measured on held-out test windows rather than derived from its inputs.

full rationale

The paper's central claim is an empirical performance ranking of PEFT variants applied to SAM on the Garrulus field-D orthomosaic. The mIoU, precision, recall, and Dice values in Table I are measured on a fixed grid of 36 held-out test windows, with training windows sampled separately, so no evaluation quantity is defined in terms of the methods' trainable parameters or fitted in a way that forces the reported ranking. The AdapterH and LoRA configurations are adopted from prior work by Houlsby et al. and Hu et al., but those citations supply the architectural definitions and the LoRA hyperparameter choice, not the forest-floor segmentation scores, and the cited results are not by the present authors. The zero-shot SAM baseline is also an independent empirical measurement. The paper explicitly acknowledges labeling inconsistencies and subjective CWD boundaries, and the 0.006 mIoU gap between AdapterH and LoRA on 36 windows is a legitimate statistical robustness concern, but label noise and small test sets affect validity and uncertainty, not circularity. No self-definitional reduction, no fitted input renamed as prediction, no self-citation load-bearing premise, and no imported uniqueness theorem appear in the manuscript. The correct finding is therefore no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper contributes an empirical comparison; there is no theoretical derivation. The free parameters are standard PEFT hyperparameters chosen from prior work rather than fitted to a model. The main unproven premises are transferability of SAM features and annotation consistency.

free parameters (3)
  • LoRA rank r = 4
    Chosen following Hu et al. [7]; no tuning on this dataset. Rank controls capacity and parameter count, so the LoRA comparison depends on this choice.
  • Adapter bottleneck dimension and scaling factor = 32, 0.1
    Same bottleneck dimension and scaling factor for both adapter variants, following prior adapter work; not optimized for the Garrulus dataset.
  • Learning rate, weight decay, epochs, batch size = 0.005, 0.001, 50, 4
    Standard training hyperparameters from prior work; no validation-based tuning is reported, and the relative ranking could shift with other settings.
assumptions (3)
  • domain assumption SAM image encoder features, with adapters or LoRA, transfer to UAV forest-floor orthomosaics.
    This is the premise of the entire setup in Section III.A; the only evidence is the authors' own test scores, so the assumption is load-bearing.
  • domain assumption The prompt-free mask decoder with per-class output tokens and learned no-mask embedding can perform semantic segmentation.
    Section III.B removes prompts and relies on learned embeddings; no external validation of this modification is shown.
  • domain assumption The field-D annotations are consistent enough to compare methods.
    Section IV.A describes ambiguous CWD definitions and Section IV.C documents labeling inconsistencies; these directly affect measured mIoU and are acknowledged by the authors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Parameter-Efficient Fine-Tuning of Vision Foundation Model for Forest Floor Segmentation from UAV Imagery." pith.science (2026). https://pith.science/paper/PDLY57R5

@misc{pith2026250508932,
  author       = {Pith},
  title        = {Pith review of: Parameter-Efficient Fine-Tuning of Vision Foundation Model for Forest Floor Segmentation from UAV Imagery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PDLY57R5}},
  note         = {Machine review of arXiv:2505.08932}
}
read the original abstract

Unmanned Aerial Vehicles (UAVs) are increasingly used for reforestation and forest monitoring, including seed dispersal in hard-to-reach terrains. However, a detailed understanding of the forest floor remains a challenge due to high natural variability, quickly changing environmental parameters, and ambiguous annotations due to unclear definitions. To address this issue, we adapt the Segment Anything Model (SAM), a vision foundation model with strong generalization capabilities, to segment forest floor objects such as tree stumps, vegetation, and woody debris. To this end, we employ parameter-efficient fine-tuning (PEFT) to fine-tune a small subset of additional model parameters while keeping the original weights fixed. We adjust SAM's mask decoder to generate masks corresponding to our dataset categories, allowing for automatic segmentation without manual prompting. Our results show that the adapter-based PEFT method achieves the highest mean intersection over union (mIoU), while Low-rank Adaptation (LoRA), with fewer parameters, offers a lightweight alternative for resource-constrained UAV platforms.

Figures

Figures reproduced from arXiv: 2505.08932 by the authors.

Figure 2
Figure 2. AdaptFormer introduces a bottleneck adapter alongside with a MLP [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. LoRA architecture with a trainable matrix product BA [7], where the [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Modified Segment Anything architecture with PEFT modules. The [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (6 more)
Figure 6
Figure 6. Figure 6: Example of randomly sampled windows within the training boundary [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]
Figure 7
Figure 7. Figure 7: Example of grid-sampled windows for the test set, sampled from the [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]
Figure 8
Figure 8. Figure 8: Labeling inconsistencies in the CWD class lead to a high number [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: CWD class labeling issues (highlighted in red), in which chopped [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]
Figure 10
Figure 10. Figure 10: Per-class mIoU, showing that performance is dominated by the [PITH_FULL_IMAGE:figures/full_fig_p005_10.png]
Figure 11
Figure 11. Figure 11: mIoU performance comparison between fine-tuning with and without [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 9 canonical work pages

  1. [1]

    UA V-assisted seeding and monitoring of reforestation sites: a review,

    I. Stamatopoulos, T. C. Le, and F. Daver, “UA V-assisted seeding and monitoring of reforestation sites: a review,”Australian Forestry, vol. 87, no. 2, pp. 90–98, 2024

    work page 2024

  2. [2]

    Factoring restoration practitioner perceptions into future design of mechanical direct seeders for native seeds,

    M. Masarei, A. L. Guzzomi, D. J. Merritt, and T. E. Erickson, “Factoring restoration practitioner perceptions into future design of mechanical direct seeders for native seeds,”Restoration Ecology, vol. 27, no. 6, pp. 1251–1262, 2019

    work page 2019

  3. [3]

    An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale,” inInternational Conference on Learning Representations, 2021

    2021

  4. [4]

    Segment Anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment Anything,” inInternational Conference on Computer Vision, 2023, pp. 4015–4026

    work page 2023

  5. [5]

    Learning Transferable Visual Models from Natural Language Supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning Transferable Visual Models from Natural Language Supervision,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 8748–8763

    work page 2021

  6. [6]

    RSPrompter: Learning to Prompt for Remote Sensing Instance Seg- mentation Based on Visual Foundation Model,

    K. Chen, C. Liu, H. Chen, H. Zhang, W. Li, Z. Zou, and Z. Shi, “RSPrompter: Learning to Prompt for Remote Sensing Instance Seg- mentation Based on Visual Foundation Model,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    work page 2024

  7. [7]

    LoRA: Low-Rank Adaptation of Large Language Models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-Rank Adaptation of Large Language Models,” inInternational Conference on Learning Representations, 2022

    2022

  8. [8]

    Parameter-Efficient Transfer Learning for NLP,

    N. Houlsby, A. Giurgiu, S. Jastrzebski, B. Morrone, Q. De Laroussilhe, A. Gesmundo, M. Attariyan, and S. Gelly, “Parameter-Efficient Transfer Learning for NLP,” inInternational Conference on Machine Learning. PMLR, 2019, pp. 2790–2799

    work page 2019

Show all 12 references
  1. [9]

    Adaptformer: Adapting Vision Transformers for Scalable Visual Recog- nition,

    S. Chen, C. Ge, Z. Tong, J. Wang, Y . Song, J. Wang, and P. Luo, “Adaptformer: Adapting Vision Transformers for Scalable Visual Recog- nition,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 664–16 678, 2022

    2022

  2. [10]

    Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation,

    J. Wu, W. Ji, Y . Liu, H. Fu, M. Xu, Y . Xu, and Y . Jin, “Medical SAM Adapter: Adapting Segment Anything Model for Medical Image Segmentation,” 2023. [Online]. Available: https: //arxiv.org/abs/2304.12620

    2023 arXiv

  3. [11]

    Segment anything model for medical image analysis: An experimental study,

    M. A. Mazurowski, H. Dong, H. Gu, J. Yang, N. Konz, and Y . Zhang, “Segment anything model for medical image analysis: An experimental study,”Medical Image Analysis, vol. 89, p. 102918, 2023

    2023

  4. [12]

    Fine-tuning vision foundation model for crack segmentation in civil infrastructures,

    K. Ge, C. Wang, Y . Guo, Y . Tang, Z. Hu, and H. Chen, “Fine-tuning vision foundation model for crack segmentation in civil infrastructures,” Construction and Building Materials, vol. 431, p. 136573, 2024. Accepted to the Novel Approaches for Precision Agriculture and Forestry...

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.