acceptodds
Withdrawn from ICLR 2027

ADATrack: Anisotropic Diffusion Aware Multi-Object Tracker for Satellite Video

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-object tracking in satellite video requires preserving identities when limited appearance cues and uncertain localization make nearby targets difficult to distinguish. This motivates a central question: how can historical evidence, current observations, and motion direction jointly shape the spatial support used for association? We propose ADATrack, an anisotropic diffusion aware tracking framework that makes this support a recurrent state for each track. Its central design couples cross-frame evidence accumulation directly to association scoring through an explicit field evolution mechanism. A learned encoder predicts an anchor and evolution parameters from track history. The retained field is warped to the predicted anchor, then updated through anisotropic diffusion, decay, and an anchor-centered source modulated by current detector responses. Warping accounts for displacement, directional diffusion spreads support preferentially along the historical motion axis, and the source injects current visual evidence. Normalized field responses sampled at detection centers define the costs for one-to-one association. After matching, the field is aligned with the observation when available and carried into the next frame. Anchor and association supervision jointly train the encoder through this recurrence. On VISO and SAT-MTB, ADATrack attains 67.1 and 60.5 HOTA, respectively, while achieving the highest IDF1 and fewest identity switches among the compared trackers.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multi-object tracking, anisotropic diffusion, recurrent spatial fields, data association, satellite video

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_TFVsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ADATrack for satellite-video multi-object tracking. It maintains a recurrent spatial field for each track and updates the field using temporal warping, anisotropic diffusion, decay, and detector objectness. Association scores are obtained by sampling the updated field at detection locations. Experiments are reported on VISO and SAT-MTB.

Critical strengths

The paper considers a relevant satellite-video tracking problem and reports experiments on two datasets.

Critical weaknesses

The methodological contribution is insufficiently justified. The recurrent dense field and anisotropic diffusion formulation appear considerably more complex than necessary, while the paper lacks a strong comparison against simpler motion-uncertainty based association models. In addition, the presentation is highly schematic and the qualitative results are difficult to interpret because the tracked targets are extremely small. Overall, I do not find the novelty and empirical evidence sufficient for ICLR.

Questions for authors

Can the authors compare ADATrack with a simpler anisotropic Gaussian/covariance-based association model using the same detector and training protocol, to demonstrate that the recurrent dense field itself is necessary?

Reviewer_FaEfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the multi-object tracking (MOT) problem in satellite videos. The authors propose ADATrack, which maintains a recursively updated "track support field" for each trajectory as the association state. On VISO and SAT-MTB, ADATrack achieves HOTA scores of 67.1 and 60.5, respectively, attaining the highest IDF1 and the fewest identity switches.

Critical strengths
  1. The core formulation exhibits genuine novelty: the "per-track recursive spatial field" is directly adopted as the source of the association cost, which is fundamentally different from Kalman point-based prediction (SORT/OC-SORT), query propagation (TrackFormer/MOTR), and diffusion-based generative motion prediction (DiffMOT).

  2. The main experimental results report substantial gains and are internally consistent across metrics: on VISO, ADATrack achieves HOTA of 67.1 compared to 62.7 for the strongest baseline MP2Net (+4.4), and on SAT-MTB, 60.5 vs. 56.0 (+4.5). Furthermore, the largest improvements appear precisely in identity-preservation metrics such as AssA, IDF1, and IDS (with IDS dropping from 437–509 to 255), which is consistent with the method's design goal of targeting association/identity and lends credibility to the internal logic of the paper.

Critical weaknesses
  1. The data split does not follow any official protocol, and the results are therefore not comparable with the literature: the paper partitions VISO and SAT-MTB into training/validation/test sets using a self-defined 7:1:2 ratio (seed 42). VISO contains only 47 videos, with roughly 9–10 sequences in the test set; all baseline numbers are re-implementations under the authors' own split, which are not comparable to the officially reported results of MP2Net, MGTrack, and others. Moreover, with such a small test set, there are no repeated runs, no error bars, and no per-sequence variance analysis, so the claimed +4.4 HOTA gain lacks statistical reliability.

  2. The evaluation of the joint detection-and-tracking baselines may be unfair: methods such as MP2Net (mask propagation) and MGTrack (motion guidance) are designed jointly for detection and tracking, and forcing them to be driven by the same frozen YOLOX-S detections may systematically weaken their original performance. The fact that MP2Net still emerges as the strongest baseline in Table 1 suggests that the extent of its degradation may even be underestimated; the paper does not report to what degree each baseline was adapted and tuned under such detector replacement.

  3. The claim of "identical detection inputs" is internally inconsistent with the large discrepancies in DetA/MOTA reported in Table 1, and this is left unexplained: if all methods consume exactly the same detection outputs, the sets of output boxes produced by different trackers should be approximately identical, and DetA should not differ by as much as 8.8 points (ByteTrack 59.5 vs. ADATrack 68.3), nor MOTA by nearly 13 points (59.8 vs. 72.6). Such a large gap in detection quality may stem from output filtering or trajectory completion strategies, but the paper does not mention this at all. Since DetA accounts for half of the weight in the HOTA framework, this unexplained factor may dominate the HOTA gains, directly undermining the causal conclusion that the gains arise from the association mechanism.

  4. The most natural baseline is missing: using the Kalman predictive density as the spatial association cost. The method can essentially be viewed as a "learnable Gaussian-mixture-style spatial cost field," yet the paper does not compare against the following simple scheme: constructing an anisotropic Gaussian association cost from the Kalman filter's predictive covariance, which grows with the number of missed frames and stretches along the direction of motion.

  5. No runtime or complexity analysis is provided: each active trajectory requires maintaining an H′×W′ field every frame and performing K convolutions (with the kernel radius m increasing with diffusion strength), so that both complexity and memory grow linearly with the number of trajectories. In the VISO scenario, small objects are dense and trajectory counts are high; the paper reports neither FPS nor GPU memory, nor efficiency comparisons against lightweight methods such as ByteTrack/OC-SORT, making it impossible to assess practical value. This omission is particularly conspicuous under the "satellite video" application narrative.

Questions for authors

See weakness.

Reviewer_m9KfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ADATrack proposes a recurrent spatial response field for association in satellite-video multi-object tracking. A frozen YOLOX-S detector provides boxes and objectness responses, while a learned encoder predicts an anchor and field-update parameters from track history. The retained field is warped to the anchor and updated through motion-aligned diffusion, decay, and an objectness-modulated local source. Normalized responses at detection centers define the costs for gated Hungarian assignment, after which observations update the stored track state. Training uses anchor and candidate-ranking losses with ground-truth-forced histories. The paper reports 67.1 HOTA on VISO and 60.5 on SAT-MTB under its own sequence splits, together with component and association ablations on VISO.

Critical strengths
  1. Meaningful evidence for the proposed components within ADATrack. Removing the recurrent field reduces VISO HOTA from 67.1 to 61.6 and AssA from 65.8 to 57.7; learned isotropic diffusion obtains 65.0 HOTA versus 67.1 for the full method. These ablations provide substantive evidence that field retention and motion direction contribute to the evaluated system.

  2. Encouraging reported results on two satellite-video datasets. On the authors' sequence splits, ADATrack reports 67.1 HOTA on VISO and 60.5 on SAT-MTB, compared with 62.7 and 56.0 for the strongest listed HOTA baseline. IDF1 and identity-switch counts also improve on both datasets. The magnitude and consistency of these reported gains support the potential usefulness of the design, subject to the comparison concerns discussed below.

Critical weaknesses

1. The paper does not convincingly show that its proposed association method is more effective than appropriately adapted existing trackers, leaving its contribution to the broader MOT community unclear. In surveying satellite-video MOT work, I found substantially different evaluation protocols even for VISO: its original paper specifies a MOT training/validation/test split, while HRTracker randomly selects 20 videos for training and five for testing. LocaLock also documents differences in matching criteria by reporting results under both near-zero and 0.4 IoU thresholds. These differences make published scores difficult to compare directly. I recognize this as a field-wide difficulty and do not expect the authors to reproduce every previous tracker. A new evaluation split can nevertheless support a fair comparison if representative baselines are appropriately adapted and evaluated under a common protocol.

The paper claims to establish such a comparison in Table 1 by using identical detector outputs, but the information provided is insufficient to assess whether the baselines were appropriately adapted. It does not explain whether and how CO-MOT, developed for generic MOT, was retrained for satellite-video MOT or how its end-to-end detection pipeline was adapted to the shared detections. MP²Net is already a satellite-video method, but its learned joint detection-and-tracking pipeline also requires explanation under this comparison. For generic detection-based trackers such as ByteTrack and OC-SORT, it is unclear whether thresholds and track-management parameters originally selected for ordinary MOT were adequately tuned for tiny satellite targets. These choices can materially affect the ranking, leaving uncertainty about whether the reported advantage reflects a stronger association method or differences in baseline adaptation.

I think that, as an ICLR submission, the paper should establish a methodological or conceptual contribution relevant to the broader MOT community, not only to researchers working on satellite-video MOT. In this paper, comparisons with existing trackers are a central part of the evidence for the proposed association representation's value. The internal ablations deserve credit for showing that ADATrack's components help within its own system, but they cannot by themselves establish what the approach adds relative to existing MOT methods. Without a sufficiently credible comparative basis, I remain unconvinced that the reported results establish a method or insight that the broader MOT community can reliably build on.

2. The paper does not establish how its recurrent spatial field resolves ambiguity between nearby targets more effectively than a simpler recurrent model of motion uncertainty. The motivation concerns identity preservation under weak appearance cues and uncertain localization. However, the detector objectness map in Eq. (4) is shared across tracks and indicates possible object locations without identifying which track an object belongs to. Distinguishing targets therefore depends on their historical states and the resulting field shapes. Spreading support along motion can tolerate longitudinal localization errors, but can also increase responses to nearby competing detections along the same axis. Whether the retained field resolves this tradeoff is not examined directly.

The component ablations provide meaningful evidence: removing the recurrent field reduces VISO HOTA from 67.1 to 61.6, while learned isotropic diffusion achieves 65.0 compared with 67.1 for the full model. These results support the usefulness of field retention and motion direction within ADATrack. However, the source-only control removes field memory, and the propagation controls vary the update within the same spatial representation. They do not establish the added value of retaining a full spatial field over a compact state that also preserves temporal information and directional uncertainty. A recurrent motion-aligned Gaussian response, using the same anchor predictions, detector evidence, gating, and track-management rules, would provide a more targeted comparison.

An analysis of association decisions should show where the full field improves over such a control and where it fails, for example under localization errors aligned with motion versus closely spaced targets competing along that axis. Changes in the score separation between the correct detection and competing detections, together with actual assignment errors or identity switches, would help connect the representation to the problem it is intended to solve. The current aggregate tables and selected TP/FP/FN visualizations do not establish this connection. Consequently, the distinct association capability provided by the recurrent field, and the conditions under which it is preferable to a simpler alternative, remain insufficiently understood.

Questions for authors
  1. Baseline adaptation and tuning — clarification. Given the fragmented evaluation practices in satellite-video MOT and the substantial effort needed to adapt different trackers, I do not expect the authors to fully adapt every comparison method. Could they focus on a few strong, representative baselines and provide a well-documented comparison under their evaluation protocol? Examples could include learned systems such as CO-MOT or MP²Net and training-free association methods such as ByteTrack or OC-SORT; these are illustrative examples, not a required list. For the selected learned systems, please clarify initialization, target-domain training, and how the common detector outputs were used while preserving each method's core components. For the selected association methods, please explain how tracking parameters were tuned using validation data.

  2. Added value over related approaches — request for evidence. Can the authors provide a fair comparison with representative methods that also address localization uncertainty and association ambiguity through motion uncertainty modeling or accumulated temporal evidence? The comparison should account for detection inputs, training, and tuning while preserving each method's core design. A simpler recurrent spatial model, such as a motion-aligned Gaussian response under shared anchor predictions, detector evidence, and matching rules, could additionally serve as a diagnostic control for the full field's added value. Such a control should retain temporal information so that it does not repeat the source-only ablation.

  3. Connection to identity ambiguity — request for evidence. On VISO, in which cases does the full field improve or worsen association relative to the relevant comparison methods? Could the authors analyze localization errors aligned with motion and closely spaced targets competing along that axis, reporting the score separation between correct and competing detections together with actual assignment errors or identity switches? This would help clarify which ambiguities the retained field resolves and where its directional spreading becomes less effective.

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

Related papers

Discussion (0)

Sign in to comment.