CoChat
Sign Up
Cited by 19Advanced Neural Network ApplicationsVideo Surveillance and Tracking MethodsVisual Attention and Saliency Detection

SAM2LONG: Enhancing SAM 2 for Long Video Segmentation with a Training-Free Memory Tree

Open access
Ding, Shuangrui, Qian, Rui, Dong, Xiaoyi, Zhang, Pan, Zang, Yuhang, Cao, Yuhang, Guo, Yuwei, Lin, Dahua +1 more·2025·2025 IEEE/CVF International Conference on Computer Vision (ICCV)
DOI 10.1109/iccv51701.2025.01264
Open this paper in CoChat — free accountOpen PDF

Abstract

The Segment Anything Model 2 (SAM 2) has emerged as a powerful foundation model for object segmentation in both images and videos, paving the way for various downstream video applications. The crucial design of SAM 2 for video segmentation is its memory module, which prompts object-aware memories from previous frames for current frame prediction. However, its greedy-selection memory design suffers from the "error accumulation" problem, where an errored or missed mask will cascade and influence the segmentation of the subsequent frames, which limits the performance of SAM 2 toward complex long-term videos. To this end, we introduce SAM2Long, an improved training-free video object segmentation strategy, which considers the segmentation uncertainty within each frame and chooses the video-level optimal results from multiple segmentation pathways in a constrained tree search manner. In practice, we maintain a fixed number of segmentation pathways throughout the video. For each frame, multiple masks are proposed based on the existing pathways, creating various candidate branches. We then select the same fixed number of branches with higher cumulative scores as the new pathways for the next frame. After processing the final frame, the pathway with the highest cumulative score is chosen as the final segmentation result. Benefiting from its heuristic search design, SAM2Long is robust toward occlusions and object reappearances, and can effectively segment and track objects for complex long-term videos. Notably, SAM2Long achieves an average improvement of 3.0 points across all 24 head-to-head comparisons, with gains of up to 5.3 points in J&F on long-term video object segmentation benchmarks such as SA-V and LVOS. The code is released at https://github.com/Mark12Ding/SAM2Long.

Verified 2026-10-22 · 1 reference not checked yet

AI snapshot

Written from the paper’s full text.

TL;DR

SAM2Long is a training-free enhancement to SAM 2 that replaces its greedy memory selection with a constrained tree search over multiple segmentation pathways. It consistently improves SAM 2 across VOS and VOT benchmarks, especially on long-term and occlusion-heavy videos.

Key findings

  • SAM2Long improves SAM 2 across nine VOS benchmarks and three VOT benchmarks, with an average gain of 3.7 points across all 12 direct comparisons and gains up to 5.3 J&F on SA-V test.
  • On SA-V, SAM2Long-L reaches 80.8 J&F on val (+4.5 over SAM 2-L) and 80.8 J&F on test (+5.3); SAM2Long-S reaches 77.7 J&F on val (+4.7 over SAM 2-S) and 78.1 J&F on test (+3.5). On LVOS v2 val, SAM2Long-L reaches 85.4 J&F (+2.4) and SAM2Long-S reaches 83.2 J&F (+3.5).
  • Ablations show P=3 pathways is best: SA-V 80.8 and LVOS 85.4, with only a 14% FPS slowdown, 8% GFlops increase, and 4% memory increase; P=4 gives no additional gains.
  • On VOT benchmarks, SAM2Long achieves LaSOT 73.9, LaSOText 60.9, and GOT-10k 81.1, compared with SAM 2’s 70.0, 56.9, and 80.7.

Methods

  • Training-free modification of SAM 2’s memory: maintain a fixed number P of memory pathways, each with its own memory bank and cumulative score. For each frame, each pathway generates three mask candidates; cumulative score adds log predicted IoU, and only the top P pathways are kept. After the final frame, the highest-scoring pathway is selected.
  • Uncertainty handling: if the maximum absolute occlusion score across pathways is below δ_conf = 2, the method selects mask candidates with distinct rounded IoU values to preserve diversity. Object-aware memory bank: frames are included only if IoU > δ_IoU = 0.3 and occlusion score > 0; memory attention is modulated with weights [w_low, w_high] = [0.95, 1.05] based on occlusion scores.
  • Evaluation: semi-supervised VOS with first-frame mask on nine VOS benchmarks (SA-V val/test, LVOS v1/v2, MOSE, VOST, PUMaVOS, DAVIS, YTVOS) and three VOT benchmarks (LaSOT, LaSOText, GOT-10k), reporting J, F, and J&F; SAM 2 baselines are re-produced under the same settings.

Limitations

  • Stated in the paper: performance is constrained by SAM 2’s inherent capacity because no learnable parameters are modified, restricting optimization potential.
  • Stated in the paper: the method primarily focuses on single-object tracking rather than carefully designed multi-object scenarios, though experiments show it performs well in both single-object and multi-object cases.
  • Stated in the paper: a failure case occurs with dynamic background changes and distracting elements, where SAM2Long either tracks the wrong object or loses the target; the authors attribute this to SAM 2’s over-reliance on fine-grained visual details and lack of semantic understanding.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/iccv51701.2025.01264

Cites

82 of 88 references verified0 citation errors · 5 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

19 of 32 cited sources checked · 7 supported · 4 flagged · accuracy not yet measured

  • Premisecitation

    Adopting a memory architecture to store object information from past frames is indispensable for accurately tracking objects in video.

    “To address the above challenges, adopting a memory architecture to store the object information from past frames is indispensable for accurately tracking objects in video [5,7,10,29,40,52,56,57,64,66,68].”
    Quote in the paperSubstantial claim10 of 10 sources checked
    • cites [5] source

      Section 3.5 (p. 10) describes retaining samples from past frames to update the target model, and the ablative study (p. 26) reports that updating it improves segmentation performance. The paper does not establish that a memory architecture is indispensable for accurate tracking; it reports a performance benefit over using the first-frame model without updates. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [7] source

      The Conclusions (p. 8) say OSVOS achieves highly accurate, temporally consistent segmentation while processing each frame independently, without explicitly modeling temporal consistency. The discussion on p. 2 likewise says its object model produces stable results even when frames are processed independently. This contradicts the claim that storing information from past frames is indispensable. (LLM, conf=0.94)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [10] source

      The 'Retrieval with Online Adaptation' section says the method adds high-confidence samples from later frames to a pool of reference samples. The ablation analysis reports a slight accuracy gain from online adaptation (+1.2% in J), but does not establish that a memory architecture is indispensable for accurate video object tracking. Page numbers are unavailable in our copy. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [29] source

      Section 3.3 (p. 7) says the method appends features from processed frames to its foreground and background feature sets, and the ablation study (Section 4.4, p. 13) reports a modest accuracy gain from those updates. But the same ablation reports 0.792 mIoU without either update, so it does not support calling storage of information from past frames indispensable. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [40] source

      The paper makes the contrary point. On p. 1, it says its method produces accurate results while processing each frame independently, without temporal information from past frames; p. 2 reports accuracy results for that method. This contradicts the claim that storing information from past frames is indispensable. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [52] source

      The Method (p. 3) says the architecture stores masks and features from past frames in memory and that using accumulated samples reduces the risk of model drift. The paper supports the value of memory for its video object segmentation approach, but does not establish that memory is indispensable for accurate tracking in general. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [56] source

      Related Work (p. 2) says the method stores temporal context from video frames implicitly in adapted network weights, and the ablation study (p. 9) reports better segmentation accuracy with online adaptation. The paper does not describe a dedicated memory architecture or establish that one is indispensable for accurate object tracking. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [57] source

      The Method section (page not supplied) says FEELVOS uses information from the first and previous frames to enable tracking, and the Ablation Study reports that previous-frame matching improves results. But the paper does not describe this as a memory architecture or establish that such an architecture is indispensable for accurate video tracking. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [64] source

      Section 3.4 (p. 4) says a rough position and size from the previous frame are enough to infer the object mask with the current RGB image, and warns that relying too heavily on the previous mask can propagate errors. Section 3 (p. 2) describes a visual and spatial modulation approach, not an indispensable memory architecture storing object information from past frames. (LLM, conf=0.84)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [66] source

      The paper describes a memory network as a feature of STMVOS (Related Work, p. 3), but does not say that such an architecture is indispensable. Instead, it describes CFBI as matching information from the first and previous frames (Related Work, p. 4) and reports that CFBI outperforms other state-of-the-art methods on three benchmarks (Introduction, p. 3). (LLM, conf=0.84)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
  • Findingcitation

    SAM2Long performs comparably to SAM 2 on shorter-duration datasets.

    “Although SAM2Long performs comparably to SAM 2 on shorter-duration datasets [47,63], this aligns with the limited long-term tracking opportunities provided by short video sequences.”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [47] Pont-Tuset, 2017, The 2017 davis challenge on video object segmentation, arXiv preprint source
    • cites [63] Xu, 2018, Youtube-vos: A large-scale video object segmentation benchmark, arXiv preprint source
  • Findingown data

    SAM2Long-S shows a 4.7-point gain over the same size SAM 2 model on SA-V val set.

    “and SAM2Long-S shows an impressive 4.7-point gain over the same size SAM 2 model on SA-V val set.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    The performance of SAM 2 gradually declines as propagation progresses into later temporal segments.

    “The performance of SAM 2 gradually declines as the propagation progresses into later temporal segments, underscoring its limitations in keeping accurate tracking over time.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    In scenarios with occlusions or reappearing objects, SAM 2 can produce mask proposals that are highly variable and uncertain.

    “However, in scenarios with occlusions or reappearing objects, it can produce mask proposals that are highly variable and uncertain.”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 24
  • May combine two claims. 39

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    SAM 2 has achieved state-of-the-art performance across various video object segmentation tasks, significantly surpassing previous methods.

    “Notably, SAM 2 [50] has achieved state-of-the-art performance across various video object segmentation tasks, significantly surpassing previous methods.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [50] Ravi, 2024, Sam 2: Segment anything in images and videos, arXiv preprint source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    XMem leverages a hierarchical memory structure for pixel-level memory reading.

    “A prominent example is XMem [11], which leverages a hierarchical memory structure for pixel-level memory reading combined.”
    Quote in the paperNot a substantial claim1 of 1 sources checked

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Cutie improves segmentation accuracy by processing pixel features at the object level to better handle complex scenarios.

    “Building on XMem's framework, Cutie [14] further improves segmentation accuracy by processing pixel features at the object level to better handle complex scenarios.”
    Quote in the paperNot a substantial claim1 of 1 sources checked

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    DAVIS2017 is a well-known benchmark dataset comprising 60 training videos and 30 validation videos, with a total of 6,298 frames.

    “DAVIS2017 [47] is a well-known benchmark dataset comprising 60 training videos and 30 validation videos, with a total of 6,298 frames.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [47] Pont-Tuset, 2017, The 2017 davis challenge on video object segmentation, arXiv preprint source

      Table 1 (p. 2) lists 60 training and 30 validation sequences for DAVIS 2017, supporting the video counts. It lists 4,219 training frames and 2,023 validation frames—6,242 combined, not 6,298—and 10,459 frames across all splits. (LLM, conf=0.94)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    LaSO-Text is an extension of LaSOT, featuring a subset of 15 categories with 150 videos.

    “LaSO-Text [23] is an extension of LaSOT, featuring a subset of 15 categories with 150 videos.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [23] source

      Section 3.2 (Data Collection) says that LaSOT has a part comprising 150 sequences from 15 object classes. Our copy does not identify that part as LaSO-Text or describe LaSO-Text as an extension of LaSOT, so it supports the figures but not their attribution to the named extension. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DAVIS2017 offers high-quality, pixel-level annotations for every frame.

    “It offers high-quality, pixel-level annotations for every frame”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [47] Pont-Tuset, 2017, The 2017 davis challenge on video object segmentation, arXiv preprint source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    DAVIS2017 is a standard resource for evaluating different VOS methods.

    “making it a standard resource for evaluating different VOS methods”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [47] Pont-Tuset, 2017, The 2017 davis challenge on video object segmentation, arXiv preprint source

  • May combine two claims.

    Findingreasoning

    SAM 2 only selects a single mask as memory, sometimes disregarding the correct one.

    “However, SAM 2 only selects a single mask as memory, sometimes disregarding the correct one.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    SAM 2's mask decoder generates multiple diverse masks, accompanied by predicted IoU scores and an occlusion score, when handling challenging and ambiguous cases.

    “Our approach is motivated by the observation that the SAM 2 mask decoder generates multiple diverse masks, accompanied by predicted IoU scores and an occlusion score when handling challenging and ambiguous cases.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    SAM2Long is completely free of additional training and does not introduce any external parameters.

    “Our improvement is completely free of additional training and does not introduce any external parameters, but simply unleashes the potential of SAM 2 itself.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    As evidenced in Figure 1(a), SAM2Long successfully resolves occlusions and re-tracks recurring dancers, where SAM 2 fails.

    “As evidenced in Figure 1(a), our approach successfully resolves occlusions and re-tracks the recurring dancers, where SAM 2 fails.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    SAM2Long achieves an average improvement of 3.7 points across all 12 direct comparisons, with gains of up to 5.3 points in J&F on long-term VOS benchmarks such as SA-V and LVOS.

    “Notably, SAM2Long achieves an average improvement of 3.7 points across all 12 direct comparisons, with gains of up to 5.3 points in J &F on long-term video object segmentation benchmarks such as SA-V and LVOS.”
    • Also stated in: § 4.2. Main Results · plain
      “The average performance across 12 experiments shows an improvement of 3.7 points in the J &F score.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Neither temporal nor spatial memory frame selection enhances performance, showing that IoU filtering remains the key factor.

    “However, as shown in Table 7, neither approach enhances performance, showing that IoU filtering remains the key factor.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    SAM2Long either mistakenly tracks the wrong shirt or completely loses the target in failure cases.

    “The model either mistakenly tracks the wrong shirt or completely loses the target.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Using P=4 shows no additional gains, suggesting that three pathways strike the optimal balance between accuracy and computational efficiency for the SAM 2 model.

    “However, using P = 4 shows no additional gains, suggesting that three pathways strike the optimal balance between accuracy and computational efficiency for the SAM 2 model.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    SAM2Long achieves competitive performance with state-of-the-art VOT methods and SAM 2 baseline.

    “SAM2Long achieves competitive performance with state-of-the-art VOT methods and SAM 2 baseline, demonstrating its strong generalization ability across various tracking scenarios.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    SAM2Long already performs exceptionally well in both single-object and multi-object cases.

    “However, our experiments demonstrate that SAM2Long already performs exceptionally well in both single-object and multi-object cases.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    Even during occlusion, SAM2Long resists tracking error and correctly resumes tracking the original dancer when she reappears.

    “Even during the occlusion, SAM2Long manages to resist the tracking error and correctly resumes tracking the original dancer when she reappears.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Findingnone given

    The image encoder of SAM 2-Large consists of 212M parameters while the total parameter of SAM 2-Large is 224M.

    “For instance, the image encoder of SAM 2-Large consists of 212M parameters while the total parameter of SAM 2-Large is 224M.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Using three pathways incurs only a 14% FPS slowdown, an 8% increase in GFlops, and a 4% rise in memory usage, while boosting performance by 4.5 points on SA-V and 2.4 points on LVOS.

    “Using three pathways incurs only a 14% FPS slowdown, an 8% increase in GFlops, and a 4% rise in memory usage, while boosting performance by 4.5 points on SA-V and 2.4 points on LVOS.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    The performance remains fairly robust across datasets when tuning the hyperparameters δ_conf, δ_IoU, and [w_low, w_high].

    “We find that the performance remains fairly robust across datasets when tuning these hyperparameters; thus, we use the same values to report results for all experiments.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    SAM2Long surpasses XMem and STCN by over 20 points on both LVOS v2 and SA-V val.

    “As a result, SAM2Long surpasses XMem and STCN by over 20 points on both LVOS v2 and SA-V val, demonstrating its superior ability to handle long-term object segmentation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    SAM 2 struggles with challenging cases involving long-term reappearing objects and confusingly similar objects.

    “However, SAM 2 struggles with challenging cases involving long-term reappearing objects and confusingly similar objects.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The mask decoder and memory module are lightweight relative to the image encoder.

    “Notably, these components are lightweight relative to the image encoder.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Lower values of δconf (e.g., 0) lead to suboptimal performance by committing to incorrect segmentations, causing error propagation.

    “Lower values (e.g., 0) lead to suboptimal performance by committing to incorrect segmentations, causing error propagation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The configuration [0.95, 1.05] achieves the best performance while increasing the modulation range decreases performance.

    “We find that the configuration of [0.95, 1.05] achieves the best performance while increasing the modulation range decreases performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    SAM2Long significantly reduces segmentation errors, showing improved accuracy and consistency in object tracking across frames.

    “SAM2Long significantly reduces segmentation errors, showing improved accuracy and consistency in object tracking across frames.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    SAM2Long achieves a 1.6% improvement in single-object sequences, a 2.9% improvement in multi-object sequences, and an overall 2.5% enhancement in performance.

    “Specifically, SAM2Long achieves a 1.6% improvement in single-object sequences, a 2.9% improvement in multi-object sequences, and an overall 2.5% enhancement in performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    SAM2Long performs exceptionally well in both single-object and multi-object cases.

    “SAM2Long performs exceptionally well in both single-object and multi-object cases.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    In the Fast & Furious movie scene, SAM2Long successfully tracks the green car, even under challenging dynamic camera movements.

    “Notably, in the Fast & Furious movie scene, SAM2Long successfully tracks the green car, even under challenging dynamic camera movements.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    SAM2Long outperforms SAM2.1 in both single-object and multi-object scenarios.

    “Our experiments demonstrate that SAM2Long outperforms SAM2.1 in both singleobject and multi-object scenarios.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Higher values of δconf (e.g., 4) do not improve performance, indicating that beyond a certain threshold, the model efficiently relies on top-scoring masks without needing additional diversity.

    “Higher values (e.g., 4) do not improve performance, indicating that beyond a certain threshold, the model efficiently relies on top-scoring masks without needing additional diversity.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Once an incorrect mask is selected into SAM 2's memory, it is uncorrectable and will mislead the segmentation of subsequent frames.

    “Once an incorrect mask is selected into memory, it is uncorrectable and will mislead the segmentation of the subsequent frames.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    These strategies help SAM 2 focus on reliable object clues and improve segmentation accuracy with negligible computational overhead.

    “These strategies help SAM 2 focus on reliable object clues and improve segmentation accuracy with negligible computational overhead.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The occlusion score o_t > 0 reliably indicates object presence and its magnitude indicates confidence.

  • May combine two claims.

    Interpretationreasoning

    SAM2Long is a training-free video object segmentation strategy that considers segmentation uncertainty within each frame and chooses video-level optimal results from multiple segmentation pathways in a constrained tree search manner.

    “To this end, we introduce SAM2Long, an improved trainingfree video object segmentation strategy, which considers the segmentation uncertainty within each frame and chooses the video-level optimal results from multiple segmentation pathways in a constrained tree search manner.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The predicted IoU scores produced by SAM 2's mask decoder are reliable enough to rank and select segmentation pathways.

  • Classified as probably not a substantial claim.

    Premisecitation

    SAM 2 incorporates a memory module that enables it to generate masklet predictions using stored memory contexts from previously observed frames.

    “Building upon the original SAM [33], SAM 2 incorporates a memory module that enables it to generate masklet predictions using stored memory contexts from previously observed frames.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [33] source

  • May combine two claims.

    Interpretationreasoning

    SAM2Long is robust toward occlusions and object reappearances and can effectively segment and track objects for complex long-term videos.

    “Benefiting from its heuristic search design, SAM2Long is robust toward occlusions and object reappearances, and can effectively segment and track objects for complex long-term videos.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Multiple Hypothesis Tracking (MHT) stands out as a robust paradigm that maintains tree search to construct and manage memory.

    “Among these, Multiple Hypothesis Tracking (MHT) [15,32,51] stands out as a robust paradigm that maintains tree search to construct and manage memory.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [15] source
    • cites [32] source
    • cites [51] source

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    SAM2Long is modified from the official codebase of SAM2 and we report the re-produced performance of SAM2 with the same official code and checkpoint.

    “SAM2Long is modified from the official codebase of SAM2 and we report the re-produced performance of SAM2 with the same official code and checkpoint.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Since SAM2Long processes the image encoder only once just as SAM 2 does, the introduction of a memory tree adds negligible computational cost while significantly enhancing SAM 2's robustness against error-prone cases.

    “Since we process the image encoder only once just as SAM 2 does, the introduction of a memory tree adds negligible computational cost while significantly enhancing SAM 2's robustness against error-prone cases.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The predicted IoU scores produced by SAM 2's mask decoder are reliable enough to rank and select segmentation pathways.

  • May combine two claims.

    Interpretationreasoning

    By employing a constrained tree memory structure with object-aware memory modulation, SAM2Long effectively mitigates error accumulation and improves robustness against occlusions, resulting in a more reliable segmentation process over extended periods.

    “By employing a constrained tree memory structure with object-aware memory modulation, SAM2Long effectively mitigates error accumulation and improves robustness against occlusions, resulting in a more reliable segmentation process over extended periods.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The predicted IoU scores produced by SAM 2's mask decoder are reliable enough to rank and select segmentation pathways.
    • Rests on an assumption: The occlusion score o_t > 0 reliably indicates object presence and its magnitude indicates confidence.

  • May combine two claims.

    Interpretationreasoning

    The constrained tree memory structure maintains multiple plausible pathways and mitigates the effects of erroneous predictions.

    “This tree-based approach maintains multiple plausible pathways and mitigates the effects of erroneous predictions.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The predicted IoU scores produced by SAM 2's mask decoder are reliable enough to rank and select segmentation pathways.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    SAM2Long's failure is attributed to SAM 2's over-reliance on fine-grained visual details and its lack of semantic understanding.

    “We attribute this to SAM 2's over-reliance on fine-grained visual details and its lack of semantic understanding.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    For a strictly fair comparison, we report both the SAM2 and SAM2Long results under the same hardware/software environment and settings.

    “We noticed a slight difference between the re-produced results and the performance reported in the SAM2 paper, but for a strictly fair comparison, we tend to report both the SAM2 and SAM2Long results under the same hardware/software environment and settings.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The enhanced diversity benefits the tree search strategy and improves performance in handling uncertain cases.

    “This enhanced diversity benefits our tree search strategy and improves performance in handling uncertain cases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    SAM2Long is a training-free enhancement to SAM 2 that alleviates its limitations in long-term video object segmentation.

    “In this paper, we introduce SAM2Long, a training-free enhancement to SAM 2 that alleviates its limitations in longterm video object segmentation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The proposed memory tree effectively boosts the model's ability to track the correct object while reducing the impact of occlusion.

    “This result demonstrates that the proposed memory tree effectively boosts the model's ability to track the correct object while reducing the impact of occlusion.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The predicted IoU scores produced by SAM 2's mask decoder are reliable enough to rank and select segmentation pathways.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    SAM2Long primarily focuses on single-object tracking rather than carefully designed multi-object scenarios.

    “Second, the method primarily focuses on single-object tracking rather than carefully designed multi-object scenarios.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    GOT-10k is a large-scale generic object tracking dataset that covers over 10,000 video sequences, spanning more than 560 object classes.

    “GOT-10k [30] is a large-scale generic object tracking dataset that covers over 10,000 video sequences, spanning more than 560 object classes.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30] source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    GOT-10k is a widely used benchmark for developing and assessing tracking models.

    “making it a widely used benchmark for developing and assessing tracking models”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30] source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    GOT-10k provides strict one-shot evaluation settings and diverse real-world tracking scenarios.

    “It provides strict one-shot evaluation settings and diverse real-world tracking scenarios”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [30] source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    LaSOT is a challenging benchmark for evaluating tracking algorithms.

    “making it a challenging benchmark for evaluating tracking algorithms”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] source

  • May combine two claims.

    Interpretationown data

    SAM2Long offers substantial improvements over SAM2, especially in handling object occlusion and reappearance, leading to better performance in long-term video segmentation tasks.

    “Overall, SAM2Long offers substantial improvements over SAM2, especially in handling object occlusion and reappearance, leading to better performance in long-term video segmentation tasks.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    LaSOT is a large-scale tracking dataset designed for long-term visual object tracking.

    “LaSOT [22] is a large-scale tracking dataset designed for long-term visual object tracking.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    LaSOT contains 1,400 videos spanning 70 object categories, with an average sequence length exceeding 2,500 frames.

    “It contains 1,400 videos spanning 70 object categories, with an average sequence length exceeding 2,500 frames”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] source

  • May combine two claims.

    Interpretationown data

    SAM2Long is robust and effective in various video segmentation tasks.

    “These results highlight SAM2Long's robustness and effectiveness in various video segmentation tasks.”
    Quote in the paperSubstantial claim
56 of 56

  • The paper relies on this without support.

    The predicted IoU scores produced by SAM 2's mask decoder are reliable enough to rank and select segmentation pathways.

    “Conditioned on the memory bank of each pathway p, the SAM 2 decoder head generates three mask candidates along with their predicted IoU scores”
    “We then calculate the cumulative scores for each possible pathway by adding the logarithm of its IoU score to the pathway's previous score”

  • The paper relies on this without support.

    The occlusion score o_t > 0 reliably indicates object presence and its magnitude indicates confidence.

    “Additionally, the mask decoder predicts a single occlusion score o t for the frame, where o t > 0 indicates object presence, o t < 0 indicates absence, and the absolute value |o t | depicts the model's confidence.”
    “We include frame i in the memory bank if it satisfies the following criteria: IoU i > δ IoU and o i > 0”
2 of 2

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Conclusion says six VOS benchmarks; evaluation says nine

    Minor

    The conclusion describes evaluations on six VOS benchmarks, while the experimental setup explicitly lists nine.

  • LVOS v1 split totals 220 videos, not 720

    Minor

    The stated training, validation, and testing splits contain 120 + 50 + 50 = 220 videos, not the stated 720.

    Reported 720 · recomputed 220

2 of 2

Cites

Flagged first, then by citations

6 of the 73 references couldn’t be matched to a known paper.

Cited by

Sorted by citations

Related

Authors
Ding, S., Qian, R., Dong, X., Zhang, P., Zang, Y., Cao, Y., Guo, Y., Lin, D., & Wang, J.
Affiliations
  • Chinese University of Hong Kong (HK)
  • Shanghai Artificial Intelligence Laboratory (CN)
ORCID
Venue
2025 IEEE/CVF International Conference on Computer Vision (ICCV)
Publisher
IEEE
Published
2025-10-19
Pages
13614-13624
DOI
10.1109/iccv51701.2025.01264
OpenAlex
W4404089453
Semantic Scholar
dcd26867448438655272a8fb12d675997e46edd4
Cited by
19
References
73
Citations by year
2026: 18 · 2025: 1
FWCI
14.68
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE/CVF International Conference on Computer Vision (ICCV) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • National Key R&D Program of China (2022ZD0161600)
  • Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Video Surveillance and Tracking Methods
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Visual Attention and Saliency Detection
Keywords
Segment Anything Model 2, video object segmentation, error accumulation, segmentation uncertainty, tree search, object tracking, occlusion robustness

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-22.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-22.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review