Pith. sign in

REVIEW 2 major objections 6 references

Turbulence-Robust Dynamic Object Segmentation with Multi-Signal Priors and SAM2 Refinement

T0 review · 2 major / 0 minor · reviewed 2026-06-29 · grok-4.3

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

Pith's one-line read A training-free pipeline fuses RAFT motion, DINOv2 semantics, ViBe background, and SAM2 refinement to segment dynamic objects in turbulence.

desk verdict A challenge report that fuses four off-the-shelf models for turbulence segmentation but gives no details on the manual fusion rules or parameters. read the letter →

arxiv 2605.29292 v1 pith:GV4JINGA submitted 2026-05-28 cs.CV

Bolian Peng , Ying Tang , Xu Liu , Long Sun , Xiaoqiang Lu This is my paper
classification cs.CV
keywords dynamicobjectsegmentationatmosphericturbulencetraining-freepipelinemulti-signalfusionSAM2refinementRAFTmotionDINOv2priorsViBebackgroundmodeling
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 introduces a system for the DOST challenge that segments moving objects in videos distorted by atmospheric turbulence without training any new model. It runs RAFT for motion, DINOv2 for semantic objectness, ViBe for background modeling, fuses their outputs through manual calibration, and refines the results with SAM2. This design addresses the problem that turbulence creates unreliable pseudo-motion and intermittent visibility, so no single cue suffices. The submitted results reach 0.425041 mIoU and 0.457206 mDice on the official leaderboard. The approach stays entirely in inference mode, making it immediately usable but dependent on the chosen pretrained components.

What carries the argument

The manually calibrated proposal fusion that integrates the four pretrained signals before SAM2 refinement.

What would settle it

Evaluating the pipeline on a new turbulence video set and finding that its mIoU falls below the score obtained by using only the RAFT motion signal.

Watch

Extended reading notes

Core claim

The central claim is that a training-free pipeline combining RAFT dense motion responses, DINOv2 semantic priors, ViBe background anomaly modeling, and SAM2 box-prompted mask refinement produces reliable segmentation masks for dynamic objects under severe atmospheric turbulence when the four signals are fused through manual calibration.

Load-bearing premise

That the four specific pretrained models can be combined through manual calibration to separate real object motion from turbulence-induced artifacts.

Editorial extensions

If this is right

  • The system runs in pure inference mode and requires no task-specific training or fine-tuning.
  • Stronger learned temporal association could raise performance on the same inputs.
  • Adaptive proposal selection might reduce reliance on manual calibration.
  • Task-specific adaptation of the components is identified as a route to further gains.

Reading between the lines

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

  • The same signal-fusion pattern could be tested on other image-degradation problems such as underwater or fogged video.
  • Replacing manual calibration with a learned selector might generalize the method beyond the DOST dataset.
  • The results suggest that ensembles of off-the-shelf models can sometimes outperform single end-to-end networks on domain shifts caused by turbulence.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 0 minor

Summary. The paper describes a training-free pipeline for the CVPR 2026 UG2+ DOST challenge that fuses RAFT motion estimates, DINOv2 semantic priors, ViBe background modeling, and SAM2 refinement via manually calibrated proposal fusion, reporting official leaderboard scores of 0.425041 mIoU and 0.457206 mDice without any task-specific training.

Significance. If reproducible, the result would demonstrate that a combination of off-the-shelf pretrained models can achieve competitive performance on turbulence-degraded dynamic segmentation without end-to-end training, providing a practical baseline for settings where labeled turbulence data are unavailable.

major comments (2)
  1. [Abstract / pipeline description] The central performance claim cannot be attributed to the described pipeline because the fusion step is described only as 'manually calibrated proposal fusion' with no thresholds, weights, decision rules, or pseudocode provided for combining RAFT, DINOv2, and ViBe signals before SAM2 refinement. This detail is load-bearing for the claim that the four-component architecture produces the reported mIoU/mDice.
  2. [Method / experiments] No ablation studies, component-wise error analysis, or sensitivity tests are reported, so the individual contributions of motion, semantic, and background priors to the final leaderboard score remain unquantified and the multi-signal design cannot be evaluated.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our technical report. We address each major point below and indicate planned revisions to improve reproducibility and evaluation of the multi-signal design.

read point-by-point responses

  1. Referee: [Abstract / pipeline description] The central performance claim cannot be attributed to the described pipeline because the fusion step is described only as 'manually calibrated proposal fusion' with no thresholds, weights, decision rules, or pseudocode provided for combining RAFT, DINOv2, and ViBe signals before SAM2 refinement. This detail is load-bearing for the claim that the four-component architecture produces the reported mIoU/mDice.

    Authors: We agree that the fusion step requires more explicit documentation to support the performance attribution. The current description is intentionally high-level due to the report format, but the fusion combines per-pixel motion magnitude from RAFT, semantic objectness scores from DINOv2, and foreground probability from ViBe using fixed thresholds and logical AND/OR rules calibrated on a small set of visual examples. In the revised manuscript we will add a dedicated subsection with the exact thresholds, weighting factors, decision logic, and pseudocode for the proposal fusion stage. revision: yes

  2. Referee: [Method / experiments] No ablation studies, component-wise error analysis, or sensitivity tests are reported, so the individual contributions of motion, semantic, and background priors to the final leaderboard score remain unquantified and the multi-signal design cannot be evaluated.

    Authors: The referee correctly notes the absence of ablations. As this is a concise technical report for a challenge submission whose primary output is the official leaderboard score, we prioritized end-to-end description over component analysis. The design rationale for using multiple signals is already stated in the text (single cues are unreliable under turbulence). We do not intend to add quantitative ablations in revision, as they would require additional experimentation outside the scope of the current report; the leaderboard result stands as the evaluation of the full pipeline. revision: no

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical leaderboard result from inference-only pipeline

full rationale

The paper reports an external leaderboard score (0.425041 mIoU, 0.457206 mDice) obtained by running a fixed combination of four pretrained models (RAFT, DINOv2, ViBe, SAM2) plus manual fusion and SAM2 refinement, all in inference mode with no task-specific training or parameter fitting to the DOST data. No equations, derivations, fitted parameters, or self-citations appear in the provided text. The result is therefore an independent empirical measurement rather than a quantity that reduces to the method's own inputs by construction. The absence of explicit fusion weights or thresholds affects reproducibility but does not create circularity under the defined criteria.

Assumptions & free parameters 1 free parameters · 1 assumptions · 0 invented entities

The central claim rests on the untested generalization of four off-the-shelf models to turbulence without adaptation and on the effectiveness of manual calibration for signal fusion; no free parameters are explicitly fitted to data in the abstract, but the manual calibration step functions as an implicit human-chosen parameter.

free parameters (1)
  • manual calibration parameters for proposal fusion
    The abstract states that proposal fusion is manually calibrated, introducing human-selected thresholds or weights that are not derived from data or theory.
assumptions (1)
  • domain assumption Pretrained models RAFT, DINOv2, ViBe, and SAM2 produce useful signals on turbulence-distorted video without task-specific fine-tuning.
    The entire pipeline is built on this generalization assumption stated in the abstract description of the training-free design.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Turbulence-Robust Dynamic Object Segmentation with Multi-Signal Priors and SAM2 Refinement." pith.science (2026). https://pith.science/paper/GV4JINGA

@misc{pith2026260529292,
  author       = {Pith},
  title        = {Pith review of: Turbulence-Robust Dynamic Object Segmentation with Multi-Signal Priors and SAM2 Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GV4JINGA}},
  note         = {Machine review of arXiv:2605.29292}
}
read the original abstract

This technical report presents our solution for the CVPR 2026 UG2+ Challenge Track 3: Dynamic Object Segmentation in Turbulence (DOST). We design a training-free multi-signal segmentation pipeline that combines pretrained motion estimation, self-supervised semantic priors, background anomaly modeling, manually calibrated proposal fusion, and SAM2-based mask refinement. The method uses RAFT for dense motion responses, DINOv2 for semantic objectness priors, ViBe for training-free background modeling, and pretrained SAM2 for box-prompt mask refinement. Instead of optimizing an end-to-end segmentation network, our system operates entirely in inference mode. This design is suitable for the DOST setting, where severe atmospheric turbulence produces pseudo-motion, blur, and intermittent target visibility, making a single motion cue unreliable. The final submitted masks are evaluated by the official leaderboard, which reports 0.425041 mIoU and 0.457206 mDice. Since no task-specific model training or fine-tuning is performed, stronger learned temporal association, adaptive proposal selection, or task-specific adaptation may further improve the system.

Figures

Figures reproduced from arXiv: 2605.29292 by the authors.

Figure 1
Figure 1. Overview of the proposed training-free DOST framework. Pretrained RAFT, DINOv2, and SAM2 models are used without DOST￾specific training. Multi-signal proposal cues are fused and refined by SAM2 box prompts with isolated-box filtering to produce the final dynamic object masks. 3.4 Skip-Frame Motion Enhancement Adjacent-frame flow can be weak for slow-moving objects. We therefore compute optical flow over a larger tem… view at source ↗
Figure 2
Figure 2. Qualitative progression of the proposed pipeline. The first row shows Person With Static And Moving Car, and the second row shows Zoomed Aeroplane. The columns show the input frame, RAFT proposal, DINO-guided proposal, skip-frame proposal, SAM2 box refinement, and the final result. CUDA 12.4, and the official SAM2 implementation. All ex￾periments are conducted on a single NVIDIA A800 80GB GPU. Since the pipeline doe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

6 extracted references · 2 canonical work pages

  1. [1]

    D. Qin, R. Saha, S. Jayasuriya, J. Ye, and N. Li. Unsupervised region-growing network for object segmentation in atmospheric turbulence.arXiv preprint arXiv:2311.03572, 2023

    work page arXiv 2023

  2. [2]

    Teed and J

    Z. Teed and J. Deng. RAFT: Recurrent all-pairs field transforms for optical flow. InEuropean Conference on Computer Vision, 2020

    2020

  3. [3]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.- W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski. DINOv2: Learning robust visual features without supe...

    2024

  4. [4]

    Barnich and M

    O. Barnich and M. Van Droogenbroeck. ViBe: A universal back- ground subtraction algorithm for video sequences.IEEE Trans- actions on Image Processing, 20(6):1709–1724, 2011

    2011

  5. [5]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Dollar, and R. Girshick. Segment anything. InIEEE/CVF In- ternational Conference on Computer Vision, 2023

    2023

  6. [6]

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Radle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Dollar, and C. Fe- ichtenhofer. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 5

    work page Pith review arXiv 2024

Pith tools

Reviewed June 29, 2026 · model on record in the stance chip above.