acceptodds
Under review as a conference paper at ICLR 2027

PRMTrack: Progressive Reliability Modeling for Satellite Video Object Tracking

Abstract

Satellite video object tracking remains challenging due to tiny object sizes, substantial aspect-ratio variations, frequent occlusions, and interference from similar objects in complex backgrounds, which often lead to template-search representation mismatch, response ambiguity, and dynamic template contamination. To address these challenges, we propose PRMTrack, a Siamese network-based progressive reliability modeling framework for satellite video object tracking that improves tracking reliability across three consecutive stages: object representation, response recalibration, and template renewal. Specifically, we develop a Multi-Scale Sparse Perception Network that strengthens structural object representations through multi-scale feature modeling and sparse interactions, alleviating feature mismatch caused by aspect-ratio variations. We further design an Adaptive Dual-Branch Response Recalibration Network that learns branch-specific Gamma mappings according to the distinct distributions of classification and centerness responses, enhancing genuine object responses while suppressing distracting peaks from similar objects. Finally, we introduce an Object-Ratio-Guided Reliable Template Renewal Mechanism that evaluates scale consistency using the area ratio between the current object and historical template and adaptively controls template renewal, thereby reducing template contamination caused by occlusions and abnormal scale estimation. Extensive experiments on SatSOT, OOTB, and SV248S demonstrate strong tracking performance, with PRMTrack achieving 74.9% precision and a 53.0% success rate on SatSOT. Source code is available at: https://anonymous.4open.science/r/Track-SVOT.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Object tracking, Reliability modeling, Multi-scale representation, Template renewal, Response recalibration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tS6wOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PRMTrack addresses unreliable information propagation in satellite video object tracking through progressive reliability modeling. It introduces three components: MSPN for multi-scale sparse feature interaction, ADRR for adaptive response recalibration, and ORTR for reliable template renewal. Experiments on SatSOT, SV248S, and OOTB demonstrate improvements over existing methods. The framework is clearly organized, although the reliability criteria and their general effectiveness require further validation.

Critical strengths
  1. The three proposed modules address representation mismatch, response ambiguity, and template contamination at different stages, providing a clear and practically motivated framework.
  2. Evaluations on three datasets and component ablation studies demonstrate promising tracking performance under various challenging conditions.
Critical weaknesses
  1. Response entropy and area-ratio constraints do not necessarily indicate correct localization or reliable observations. Their effectiveness in identifying unreliable predictions remains insufficiently established.
  2. Unclear novelty and progressive design: The technical advantages over existing feature interaction, response calibration, and template update methods are not sufficiently clarified. Further evidence is needed to establish the complementary effects of the three modules.
Questions for authors
  1. How do the authors validate that response entropy and area-ratio constraints reliably distinguish correct predictions from unreliable observations? How sensitive is ORTR to its predefined thresholds?
  2. What are the specific technical differences between the proposed modules and the most closely related methods? What evidence demonstrates that their progressive integration provides benefits beyond combining the individual modules?
Reviewer_yGVeOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the key challenges in Satellite Video Object Tracking (SVOT), where targets typically occupy very few pixels and suffer from dramatic aspect-ratio changes, background clutter, and severe occlusions. The authors present PRMTrack, a Siamese-based progressive reliability modeling tracker structured across three consecutive stages, including Multi-Scale Sparse Perception Network (MSPN), Adaptive Dual-Branch Response Recalibration Network (ADRR) and Object-Ratio-Guided Reliable Template Renewal (ORTR). Validation is conducted across three satellite benchmarks (SatSOT, SV248S, and OOTB), where PRMTrack demonstrates top-ranking precision and success rates (e.g., 74.9% precision and 53.0% success rate on SatSOT).

Critical strengths

The submitted manuscript demonstrates strengths in the formulation of algorithmic problems, analysis of model architecture, and experimental validation. The pipeline logically disentangles and tackles the three cascading failure modes in satellite tracking (representation mismatch to response ambiguity to template contamination). Rather than relying solely on generic transformer attention, the progressive constraints offer clear domain-specific value for satellite video scenarios. Moreover, the dynamic Gamma transformation dynamically adjusts the contrast between peak and background based on response entropy. It is theoretically well-aligned with the distinct functional roles of classification (concave mapping) and centerness (convex mapping), remains fully differentiable, and adds minimal computational overhead. The tracker consistently achieves state-of-the-art results across SatSOT, SV248S, and OOTB. Crucially, the ablation study in Table 4 directly validates the intended functional components against specific challenging attributes.

Critical weaknesses

First, the paper lacks experiments on inference efficiency, which is a critical metric in the field of video tracking. The paper provides no evaluation of tracking speed (FPS), computational complexity (FLOPs), or parameter count compared to competing methods (e.g., SiamTITP, HTTrack, SeqTrack). Second, the paper shows ambiguities in MSPN Token Flow and Base Architecture. Section 3 introduces Z0 (initial template) and Zd (dynamic template), but Eq. (1) applies pooling to a single undefined Z. The exact mechanism for token interaction between Z0 and Zd remains unclear. Furthermore, Eq. (1) utilizes 1D pooling, yet it is not explained how the 2D spatial feature map is serialized or flattened prior to pooling. Moreover, the paper does not explicitly state the exact baseline architecture in Section 3 (only mentioning loss terms from SiamCAR in Section 3.4). In Appendix Eq. (13), the authors mention the threshold obtained via numerical analysis, yet the derivation process is not clearly presented. Satellite targets can experience legitimate perspective compression or abrupt scale variations. The paper lacks a parameter sensitivity analysis for these bounds, and does not address how the tracker recovers if an erroneous template is accepted within this valid range. Regarding the model training method, given the Top-K truncation and conditional branches, the authors need to elaborate more clearly on the gradient propagation process of the model.

Questions for authors
  1. Satellite video tracking often demands near-real-time performance for practical deployability. Could you provide a comprehensive comparison against representative state-of-the-art trackers (e.g., SiamTITP, HTTrack, SeqTrack) regarding Frames Per Second (FPS), FLOPs / GFLOPs, and Parameter count?

  2. The paper shows ambiguities in MSPN Token Flow and Base Architecture. Section 3 introduces Z0 (initial template) and Zd (dynamic template), but Eq. (1) applies pooling to a single undefined Z. The exact mechanism for token interaction between Z0 and Zd remains unclear. Furthermore, Eq. (1) utilizes 1D pooling (AvgPoolks), yet it is not explained how the 2D spatial feature map is serialized or flattened prior to pooling. Moreover, the paper does not explicitly state the exact baseline architecture in Section 3. Could the authors formally specify the exact backbone and cross-correlation architecture of the “Baseline” tracker?

  3. In Appendix B, the area-ratio bounds [0.7,1.5] are stated as being based on “statistical analysis”. Could you present the concrete empirical derivation, distribution histograms, or validation data that led to these exact numbers? If an inaccurate or drifted bounding box happens to fall within this [0.7,1.5] ratio and is updated into the dynamic template, how does the model recover from this historical template corruption in subsequent frames?

  4. In Eq. (4), a hard Top-K selection zeroes out connections outside the top indices. Is this non-differentiable discrete masking operator bypassed during backward propagation, or does gradient backpropagation flow exclusively through the active top-K paths? Please elaborate on how the learnable weights receive stable gradient updates without getting trapped in local sub-optimal topologies.

  5. The paper still lacks necessary ablation experiments to verify the performance gains brought by each component. For example, in MSPN, what is the impact of different K values in Top-K selection compared to standard dense self-attention? Furthermore, how does the multi-scale pooling branch compare against a standard single-scale or 2D strip-pooling baseline? In ADRR, how much does the learned entropy-driven parameterization outperform a fixed, manually tuned Gamma transformation, could authors provide an ablation showing the entropy distribution and the trajectory of the learned values across typical sequences?

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

The manuscript investigates the satellite video object tracking (SVOT) problem, which is a subfield of visual object tracking characterized by extremely small object pixels, extreme aspect ratio variations, frequent occlusions, and easy confusion with background objects with similar appearances. The author proposes a tracking framework PRMTrack based on Siamese twin networks, and decomposes the SVOT robustness problem into progressive reliability modeling across three consecutive stages: (1) Enhancing representation through a Multi-Scale Sparse Perception Network (MSPN). (2) Correcting the response map using an Adaptive Dual-Branch Response Recalibration Network (ADRR). (3) Object-Ratio-Guided Reliable Template Renewal Mechanism (ORTR) guided by area ratio for gated template update. SOTA results were reported in experiments on three SVOT benchmark datasets (SatSOT, SV248S, OOTB).

Critical strengths
  1. The manuscript accurately identified three types of failure modes unique to satellite tracking, including template search feature mismatch under extreme aspect ratio variations, response ambiguity peaks caused by similar objects, and dynamic template contamination under occlusion. Three modules have been designed to address these issues one by one. The narrative logic from the generation of unreliable information to the dissemination of unreliable information is coherent and easy to understand.
  2. The author conducts evaluations on three standard SVOT datasets, covering two annotation forms: horizontal box annotation (SatSOT, SV248S) and rotation box annotation (OOTB). The author conducts fine-grained analysis on 11 challenge attributes, and qualitative visualization also supports the claimed benefits of each module. Although the ablation experiments in Table 4 are limited, they do indicate that all three modules contribute positively to their respective target attributes.
Critical weaknesses
  1. The three modules proposed in the manuscript use completely independent reliability signals. MSPN uses attention sparsity, ADRR uses response entropy, and ORTR uses bounding box area ratio. There is no unified reliability measurement transmission among the three, nor is there any discussion on how these three signals can coordinate in case of conflict. The narrative of 'progressive' is more of a textual packaging rather than a shared computational mechanism.
  2. The manuscript does not provide the specific value of K in TopK sparse selection, only indicating that two branches used different sparsity levels. The authors did not conduct any sensitivity analysis on the area ratio threshold of ORTR and the kernel size for multi-scale pooling selection.
  3. The manuscript does not report the inference speed, parameter count, and FLOPs of the proposed method, making it impossible to make reliable horizontal comparisons with other methods.
  4. Table 1 and Table 2 can indicate which ones are SOT methods and which ones are SVOT methods (such as HTTrack, SiamTITP, etc.) for clearer expression.
  5. The layout of Table 3 is too long, which is not conducive to readers' reading.
Questions for authors
  1. Has the author considered unifying attention sparsity, response entropy, and area ratio into a reliability score for end-to-end learning? How does the system coordinate when the three signals give inconsistent judgments?
  2. What are the K values used for the two TopK sparse branches? How are the area ratio threshold of ORTR [0.7, 1.5] and the kernel size of multi-scale pooling {3, 5, 7} determined? Can sensitivity analysis of these key hyperparameters be supplemented?
  3. What are the inference speed (FPS), parameter count, and FLOPs of PRMTrack, and can they reach real-time levels?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.