{"id":"c3416c21-e507-47df-b365-1a287c24ed9d","arxiv_id":"2505.03507","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GDSTrack combines modality-guided dynamic graph fusion with temporal graph-informed diffusion to train a self-supervised RGB-T tracker, outperforming prior self-supervised methods on multiple benchmarks.","lead":"This paper presents GDSTrack, a self-supervised RGB-T tracker that fuses visible and thermal features with dynamic graph attention and uses a diffusion model to suppress noise in automatically generated pseudo-labels. The authors report state-of-the-art results on four RGB-T tracking benchmarks, reducing reliance on manual annotations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LasHeR evaluation lacks a specified train/test split, so the reported 6.1% PR gain over S2OTFormer is unverifiable and may reflect in-distribution overfitting.","rationale":"The reader's weakest_assumption concerns the diffuser-noise design in Eq. (20): the neighboring-frame feature f_{x1}^{S1} may not be a true distractor if the object is static or the offset is negligible. This is a legitimate theoretical concern, but it does not directly invalidate the numerical claim: even if the diffusion mechanism is motivated imperfectly, the reported gains could still be real. The train/test split issue is more load-bearing because it threatens the validity of the reported numbers themselves. The paper trains on LasHeR and reports LasHeR results without ever describing the split. This is an explicit omission in a key part of the evaluation. If the model trains and tests on the same LasHeR sequences, the 6.1% PR advantage over S2OTFormer is not a valid generalization result. The reader's rationale lists \"unclear train/test split for LasHeR\" as one of the gaps, but the reader's weakest_assumption selected a different point. I agree the split issue is the more fundamental threat to the central claim. Therefore, the paper should be accepted only after the authors clarify the exact protocol, ideally by releasing a specific split or adding the standard evaluation description. This does not require rejecting the method out of hand, but it does require a concrete verification step before the SOTA claim is accepted.","tokens_in":13487,"tokens_out":7113,"duration_ms":65475,"concrete_test":"Inspect the released code (https://github.com/LiShenglana/GDSTrack) to identify the exact LasHeR sequences used for training and for evaluation. If the two sets overlap or if no split is defined, the LasHeR row in Table 1 is invalid and the claim must be revised. If a disjoint split exists, re-run the S2OTFormer comparison under identical training and evaluation conditions to confirm the 6.1% gain.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that GDSTrack surpasses S2OTFormer on LasHeR by 6.1% PR rests on the evaluation in Table 1. Section 4 (Implementation Details) states only \"We use LasHeR [Li et al., 2021] as our training dataset\" and never specifies a train/test split for the LasHeR rows. If the model is trained and evaluated on the same sequences, the LasHeR numbers reflect in-distribution fitting rather than generalization, invalidating the comparison to S2OTFormer. Even if a split exists, its absence in the text and the lack of any protocol description make the result irreproducible. This is a load-bearing gap because 6.1% PR is one of the three headline improvements, and the same unstated split could affect the cross-dataset results if the training set is not properly disjoint from the evaluation sets.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GDSTrack, a self-supervised RGB-T tracking method that combines a Modality-guided Dynamic Graph Fusion (MDGF) module with a Temporal Graph-Informed Diffusion (TGID) module. MDGF dynamically generates an adjacency matrix via the AMG module and uses it to guide graph attention for fusing RGB and thermal features. TGID treats the MDGF fusion output of a neighboring frame as a distractor-noise term in a DDIM-based diffusion process, aiming to improve robustness to similar-object interference. The model is trained using pseudo-labels generated from optical flow, and evaluated on GTOT, RGBT234, LasHeR, and VTUAV, reporting improvements over prior self-supervised trackers, with code released.","tokens_in":13688,"tokens_out":7432,"duration_ms":63614,"significance":"If the reported results are reproducible, GDSTrack would be a meaningful advance for self-supervised multi-modal tracking, where only a few methods exist. The combination of graph-based fusion and diffusion-based denoising is novel in this context, and the substantial margins on LasHeR and VTUAV suggest that the architecture addresses a real gap. The release of source code is a strength. However, the evaluation protocol for LasHeR is incompletely specified, the TopK hyperparameter appears to be tuned on test sets, and the TGID module's contribution is small relative to the paper's claims; these issues need to be resolved before the results can be fully trusted.","major_comments":[{"comment":"The evaluation protocol for LasHeR is not specified. The text says 'We use LasHeR as our training dataset' but gives no train/test split, so the reported PR/NPR/SR on LasHeR are unverifiable. If training and evaluation use the same sequences, the 6.1% PR gain over S2OTFormer would reflect in-distribution fitting rather than generalization. Please state the exact split (e.g., number of training/test sequences, random seed) and, if available, the standard protocol, or clarify the evaluation procedure.","section":"Section 4 (Implementation Details), Table 1"},{"comment":"The TopK hyperparameter in the AMG is selected based on test-set performance on GTOT and RGBT234 ('When K is set to 256, the model performs the best'), which is a post-hoc choice that inflates the reported numbers. In addition, the threshold θ in Eq. (14) is never given a value, so the AMG module is not fully reproducible. Please either use a validation split for model selection or explicitly acknowledge test-set tuning, and report θ and all other hyperparameters.","section":"Section 4.2, Table 5 and Eq. (14)"},{"comment":"The ablation of the TGID module shows gains of only 0.2% SR on RGBT234 and 0.2% PR/SR on LasHeR (Table 3), which are within the range of run-to-run variation, and the diffusion component ablation in Table 6 is run with K=5 rather than the final K=256. These results do not support the claim that TGID 'significantly' improves robustness to distractor noise. Please report the ablation under the same configuration as the full model and provide an analysis of statistical significance.","section":"Section 4.2, Tables 3 and 6"},{"comment":"The design assumes that the fused feature f_{x1}^{S1} from the neighboring frame is offset from the current object location and therefore acts as a similar-object distractor. If the object is static or the offset is small, this term is instead a corrupted version of the target, and the diffusion model may learn incorrect noise statistics. Please validate this assumption by reporting the distribution of inter-frame object displacement in the training data, or by an experiment with stationary objects.","section":"Section 3.3, Eq. (20)"}],"minor_comments":[{"comment":"The string 'VTUA V' should be 'VTUAV' (Abstract, Section 4.1, Table 1). Also, 'Y ong Zhou' in the author list should be 'Yong Zhou'.","section":"Abstract and throughout"},{"comment":"The statement that M is shifted to [0,1] is not justified because S1 is an unscaled dot product; please clarify the normalization or define the range precisely.","section":"Section 3.2, Eq. (13)"},{"comment":"The number of diffusion timesteps, the noise schedule, and the value of β in Eq. (20) are not reported; please add these to the implementation details.","section":"Section 4 (Implementation Details)"},{"comment":"The header 'Naive DM condition distractor' is ambiguous; please label the rows clearly (e.g., with check marks) to indicate which components are enabled.","section":"Section 4.2, Table 6"},{"comment":"The AMG ablation shows a 0.4% SR drop on RGBT234 versus the identity matrix; the text explains this as a 'little decrease,' but the PR gain is 1.1%; please discuss this trade-off.","section":"Section 4.2, Table 4"},{"comment":"The pseudo-label generation method is not described; the paper references USOT but does not explain how optical flow is applied to thermal images and how the two modalities' pseudo-labels are merged; please provide details or a specific citation to a prior RGB-T pseudo-label procedure.","section":"Section 4 (Implementation Details)"},{"comment":"All experiments appear to be single runs without error bars or statistical tests; given the small performance gaps on some benchmarks (e.g., 0.8% SR on RGBT234), please report multiple runs with mean and standard deviation, or justify the absence.","section":"Section 4.1, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's main claim hinges on a reproducible evaluation protocol. The LasHeR training/test split must be clarified. The authors should also be encouraged to release the exact split and pseudo-label generation code. The self-citation pattern is limited to baseline methods and is not problematic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GDSTrack is a plausible incremental contribution to self-supervised RGB-T tracking. The new part is the combination: a learned adjacency mask (AMG) guiding graph attention fusion, and a DDIM denoiser that uses a neighboring frame's fused feature as a 'distractor' to make the model robust to pseudo-label noise. Each component is borrowed, but the specific integration is new and clearly described. The authors also did a fair thing by retraining ViPT, TBSI, and AFter with the same pseudo-labels, making the comparisons more honest than simply citing published numbers. The ablation structure is clear, and code is promised. The soft spots are concentrated in the evaluation. The largest is the LasHeR protocol. The paper says 'We use LasHeR as our training dataset' and then reports LasHeR results in Table 1, without ever specifying how the 1,224 videos are split into training and test sets. If the same sequences are used for training and testing, the 6.1% PR gain over S2OTFormer is meaningless. If a split exists, the paper must state the number of training and test sequences and the splitting rule. This is not a minor omission; it is central to the main claim. Second, the TopK hyperparameter (k=256) is chosen after looking at test-set performance on GTOT and RGBT234 (Table 5). That is test-set overfitting, and it inflates the reported numbers. Third, the TGID module's benefit is small—0.2% SR on RGBT234 and LasHeR—so the paper should not overclaim its contribution. No error bars are given, which matters when the differences are a few tenths of a percent. The conceptual worry about the distractor noise in Sec 3.3 is less damaging than it first appears: in most tracking videos, object motion creates a real offset, so the neighboring fusion result does act as a distractor. But the paper does not analyze static-object cases, where the 'distractor' is just a corrupted target. That is a minor omission. Who is this for? Researchers working on RGB-T tracking or self-supervised tracking. It deserves a serious referee, because the architecture is coherent and the problem is real. I would ask for a rewritten experiments section with a clear data split, error bars, and hyperparameter selection on a validation set. If the LasHeR split cannot be clarified, the paper still makes a moderate contribution on RGBT234 and VTUAV, but not with the 6.1% headline. The citation pattern is fine; self-citations are mostly to baseline methods and previous work.","headline":"Solid incremental architecture, but the evaluation reporting—especially the LasHeR train/test split—needs fixing before the headline numbers can be believed.","tokens_in":709,"tokens_out":1126,"would_cite":false,"duration_ms":33426,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Self-supervised RGB-T tracking can beat supervised baselines by fusing modalities with a dynamic graph and denoising with a temporal diffusion model that treats neighboring frames as distractors.","keywords":["self-supervised tracking","RGB-T tracking","dynamic graph fusion","temporal diffusion","pseudo-label noise","multi-modal fusion","graph attention network","object tracking"],"falsifier":"Run GDSTrack on a controlled video where the target is stationary for many frames (or include a static camera sequence with no object motion) and compare the full model against the same model trained with only Gaussian noise in Eq. (20): if the neighboring-frame distractor term $d_t$ is helping, its benefit should vanish or reverse when there is no positional offset to make it look like a different, similar object. More directly, inspect the noise residual in Eq. (20): when the object does not move, the distractor term equals a shifted copy of the true target features, and the denoiser's learned residual should be measurably different from that learned on moving-object sequences.","tokens_in":13312,"feed_emoji":"🎯","tokens_out":8956,"duration_ms":75717,"temperature":0.7,"pith_summary":"This paper tries to show that a self-supervised RGB-T tracker can match or beat fully supervised approaches by attacking the two failure modes of pseudo-label training: fusion that drifts onto background regions, and confusion caused by objects that look like the target. The proposed GDSTrack builds a dynamic graph over RGB and thermal features, using a similarity-based adjacency matrix to keep attention on coherent object regions, then runs a diffusion model that treats the fused features of a neighboring frame as synthetic 'distractor noise' so the tracker learns to ignore similar objects. On RGBT234, LasHeR, and VTUAV, the method reports higher precision and success rates than the previous best self-supervised tracker, with the largest gain on the UAV benchmark. The practical stakes are high-quality RGB-T tracking without manual bounding-box annotations, which matters for large-scale and always-changing scenes.","feed_headline":"New self-supervised RGB-T tracker beats prior best on three benchmarks","feed_subtitle":"It treats neighboring-frame fused features as diffusion noise, making the tracker robust to similar-object distractors.","key_machinery":"The engine of the paper is the interplay of two modules. The Adjacency Matrix Generator (AMG) builds a dynamic adjacency matrix $A$ from the sum of scaled dot-product similarity and cosine similarity between concatenated RGB and thermal feature sequences, thresholds it, applies softmax, and keeps the Top-K connections symmetrized; this matrix guides a two-layer Graph Attention Network (GAT) that treats every spatial location in both modalities as a graph node, producing first-layer ($f^{l1}$) and fused ($f^{S1}$) features. The Temporal Graph-Informed Diffusion (TGID) module takes $f^{S1}$ of the current frame as the diffusion input $x_0$, injects the neighboring frame's $f^{S1}$ as a non-Gaussian 'distractor' additive term $d_t$ alongside standard Gaussian noise in the forward process, and denoises with a U-Net whose intermediate layers are conditioned on the first-layer graph features via a ConFuse residual-addition module. That conditioning is what prevents information loss when the diffusion model is asked to remove the distractor-like noise.","core_discovery":"On its own terms, the paper's central claim is that both dominant errors of pseudo-label-supervised RGB-T tracking can be modeled and corrected by a two-module design. The Modality-guided Dynamic Graph Fusion (MDGF) module computes a dynamic adjacency matrix from cross-modal cosine similarity and scaled dot-product attention, then uses it to guide a graph attention network so that fusion concentrates on the object's coherent regions even when the pseudo-label includes background. The Temporal Graph-Informed Diffusion (TGID) module feeds the MDGF output of the current frame into a DDIM denoising process whose added noise combines Gaussian noise with the MDGF output of a neighboring frame; because that neighboring-frame feature is offset from the target by object motion, it acts as a realistic similar-object distractor. The paper reports that GDSTrack surpasses the previous best self-supervised method S2OTFormer by 2.5% precision on RGBT234, 6.1% on LasHeR, and 15.6% on VTUAV, while remaining comparable in speed.","pith_inferences":["The distractor-noise trick is not tied to RGB-T tracking: any self-supervised task with temporal correspondences (e.g., video object segmentation, multi-object tracking) could reuse a neighboring-frame feature map as a cheap source of semantically meaningful noise for a denoiser.","The AMG adjacency matrix could be swapped into other multi-modal fusion problems (e.g., audio-visual or depth-RGB fusion) as a drop-in attention mask, since it only needs two feature tensors and no labels.","A natural test of the paper's mechanism would be to vary the frame gap used for the neighboring-frame distractor: larger gaps should create stronger 'similar object' offsets, but also increase the chance that the same object appears in very different pose, so performance should peak at some intermediate gap.","An open question the paper leaves implicit is whether the denoiser treats the distractor as clutter or as an extra appearance cue; ablating the distractor under varying motion could separate these two interpretations."],"forward_implications":["Under pseudo-label supervision, GDSTrack reports higher precision and success than the previous best self-supervised tracker on RGBT234, LasHeR, and VTUAV, with the largest improvement on the UAV set.","The MDGF module alone improves fusion under noisy pseudo-labels (roughly 2-5 points of precision across datasets compared to the baseline), showing that dynamic adjacency guided by cross-modal similarity recovers object-relevant fusion without ground truth.","The TGID module adds further gains on top of MDGF (up to 2 points of precision on VTUAV), indicating that modeling similar-object noise as temporal distractor noise makes the tracker more robust to interference.","Because the tracker trains with pseudo-labels generated from multi-modal optical flow, it removes the need for manual bounding-box annotations in the training pipeline."],"supporting_citations":[{"why":"S2OTFormer, the previous best self-supervised RGB-T tracker that serves as the primary SOTA baseline for all comparisons.","marker":"[Li et al., 2024]"},{"why":"USOT, whose optical-flow and dynamic-programming pipeline GDSTrack uses to generate pseudo-labels and confidence scores from unlabeled RGB-T video.","marker":"[Zheng et al., 2021]"},{"why":"GMMT, which first models label noise in multi-modal tracking with a generative diffusion model; TGID extends this idea to pseudo-label noise.","marker":"[Tang et al., 2024]"},{"why":"DDIM, the denoising diffusion implicit model used as the core of the TGID denoising process.","marker":"[Song et al., 2020]"},{"why":"GAT, the graph attention network that MDGF guides with the dynamically generated adjacency matrix to fuse RGB and thermal nodes.","marker":"[Velickovic et al., 2017]"},{"why":"ViPT, a prompt-based RGB-T tracker that is re-trained under the same pseudo-labels (as ViPT*) to demonstrate the performance drop the method overcomes.","marker":"[Zhu et al., 2023]"}],"fun_headline_variants":["Graph fusion + temporal diffusion for self-supervised RGB-T tracking","GDSTrack: dynamic graph fusion + diffusion noise for tracking","Self-supervised RGB-T tracker uses diffusion to beat SOTA","Modality-guided graphs + temporal diffusion improve tracking","Beats prior best on three RGB-T benchmarks via diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fused feature map of the neighboring frame really is a good stand-in for a similar-object distractor: if the tracked object barely moves or the offset regions do not look like a plausible target, the 'distractor noise' injected into the diffusion process is really a corrupted version of the target itself, and the denoiser could learn to trust incorrect noise statistics.","fun_headline_variants_meta":{"raw":{"variants":["Graph fusion + temporal diffusion for self-supervised RGB-T tracking","GDSTrack: dynamic graph fusion + diffusion noise for tracking","Self-supervised RGB-T tracker uses diffusion to beat SOTA","Modality-guided graphs + temporal diffusion improve tracking","Beats prior best on three RGB-T benchmarks via diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001061,"raw_usage":{"total_tokens":4469,"prompt_tokens":982,"completion_tokens":3487,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":3415}},"tokens_in":598,"tokens_out":3487,"duration_ms":21272,"temperature":1.0,"reasoning_tokens":3415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:49:31.323115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GDSTrack on a controlled video where the target is stationary for many frames (or include a static camera sequence with no object motion) and compare the full model against the same model trained with only Gaussian noise in Eq. (20): if the neighboring-frame distractor term $d_t$ is helping, its benefit should vanish or reverse when there is no positional offset to make it look like a different, similar object. More directly, inspect the noise residual in Eq. (20): when the object does not move, the distractor term equals a shifted copy of the true target features, and the denoiser's learned residual should be measurably different from that learned on moving-object sequences.","supporting_citations":[{"cited_title":"Learning to track objects from un- labeled videos","cited_arxiv_id":null,"evidence_quote":"USOT, whose optical-flow and dynamic-programming pipeline GDSTrack uses to generate pseudo-labels and confidence scores from unlabeled RGB-T video."},{"cited_title":"Generative-based fusion mechanism for multi-modal tracking","cited_arxiv_id":null,"evidence_quote":"GMMT, which first models label noise in multi-modal tracking with a generative diffusion model; TGID extends this idea to pseudo-label noise."},{"cited_title":"Graph attention networks.stat, 1050(20):10–48550,","cited_arxiv_id":null,"evidence_quote":"GAT, the graph attention network that MDGF guides with the dynamically generated adjacency matrix to fuse RGB and thermal nodes."},{"cited_title":"Visual prompt multi-modal track- ing","cited_arxiv_id":null,"evidence_quote":"ViPT, a prompt-based RGB-T tracker that is re-trained under the same pseudo-labels (as ViPT*) to demonstrate the performance drop the method overcomes."}],"review_version":1}