acceptodds
Withdrawn from ICLR 2027

CISTA-Net: Counting Matters for Closely-Spaced Infrared Small Target Unmixing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Closely-spaced infrared small target unmixing is an ill-conditioned inverse problem in which multiple sparse source configurations can produce highly similar observations under severe point-spread-function overlap. Existing deep unfolding methods adapt reconstruction parameters primarily from local or intermediate reconstruction features, but do not explicitly exploit scene-level source cardinality when deciding how aggressively sparse responses should be retained or suppressed. We investigate whether estimated target cardinality can serve as a global structural cue for iterative reconstruction, and instantiate this idea in CISTA-Net, a count-guided iterative shrinkage-thresholding network. CISTA-Net introduces a count-prior guided dynamic threshold generator (CP-DTG) that conditions stage-wise shrinkage on both local reconstruction features and a predicted count distribution. It further incorporates multi-stage image selective fusion (MSIF) to recover complementary intermediate responses and count-guided post-processing (CGPP) for count-consistent candidate selection. On CSIST-100K, CISTA-Net improves CSO-mAP over DISTA-Net from 46.74 to 47.28, 69.58 to 70.25, and 72.84 to 73.06 at sampling ratios \(c=3,5,7\), respectively, with the clearest gains under dense and moderately precise localization settings. These results suggest that explicit cardinality information provides complementary global guidance to locally adaptive deep unfolding.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Deep Unfolding, Sparse Recovery, Inverse Problems, Count-Conditioned Learning, Infrared Small Target Unmixing

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_zR8yOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This work addresses the problem of Closely-spaced infrared small target (CSIST) and proposes a network CISTA-Net based on deep unfolding. The core idea is to treat the number of targets as a global prior and directly apply it to control the shrinkage threshold in ISTA. It achieves "counting condition shrinkage" through CP-DTG, MSIF, and CGPP. On the CSIST-100K dataset, this method has achieved a certain improvement of +0.54 mAP compared to DISTA-Net.

Critical strengths

The paper presents a well-structured and technically coherent extension of deep unfolding for CSIST unmixing by explicitly incorporating target cardinality into the iterative reconstruction process. Its core contribution—count-guided dynamic thresholding—introduces a principled way to couple global scene-level information with local reconstruction features, improving the adaptivity of the shrinkage operator within the ISTA framework.

Critical weaknesses
  1. Limited technical novelty. The core idea of using target count to modulate the shrinkage threshold is essentially a form of feature conditioning rather than a fundamentally new paradigm or theoretical advancement. The overall framework still follows the conventional ISTA-based deep unfolding pipeline, and the proposed modification does not significantly deviate from existing formulations in terms of optimization principles.
  2. The paper lacks comparisons with several relevant and recent deep unfolding methods, such as RPCANet and DeRefNet. In addition, the method is not evaluated on more challenging or widely used benchmarks like SeqCSIST. This limits the credibility and generalizability of the reported performance gains.
  3. Compared to DISTA-Net, the proposed method introduces additional model complexity but yields only limited performance gains. Specifically, the improvements in mAP across different sampling settings (c = 3, 5, 7) are only +0.54, +0.67, and +0.22, respectively. Such marginal gains do not convincingly justify the increased model size and added architectural components.
Questions for authors

From Table 1, compared to DISTA-Net, AP-05 and AP-10 have either decreased or remained unchanged. Could this mean that by sacrificing stricter threshold performance, the overall performance has been improved? Because for more lenient thresholds, such as AP-25, the improvement is easier to achieve. Additionally, referring to weaknesses.

Reviewer_FBKEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the unmixing of closely-spaced infrared small targets whose point-spread functions overlap. CISTA-Net predicts a target-count distribution and combines it with local reconstruction features through CP-DTG to generate stage-wise shrinkage thresholds. It also uses MSIF to fuse intermediate reconstructions and CGPP to select candidates according to the predicted count. On CSIST-100K, the reported CSO-mAP is 47.28, 70.25, and 73.06 at sampling-grid settings c=3,5,7, respectively, improving over DISTA-Net by 0.54, 0.67, and 0.22 percentage points.

Critical strengths

Using the full count distribution as a soft condition for stage-wise threshold generation is a design choice worth investigating. The model retains explicit data-consistency updates and reports small CSO-mAP improvements at all three sampling-grid settings. These results provide preliminary evidence that the design may be useful, although its additional value over existing count-guided methods needs clearer experimental support.

Critical weaknesses
  1. The paper does not sufficiently establish its additional contribution over existing count-guided methods. The introduction characterizes previous work as mainly using count information at the feature level and presents direct threshold conditioning as the main distinction. However, Equation (13) of the cited DISTA-Net++ already uses count information to modulate dynamic thresholds. Using the full count distribution and adding MSIF may be useful, but the experiments do not establish what additional insight these changes provide. The small gains over DISTA-Net do not resolve this concern, and I do not find the additional contribution sufficient for acceptance.

  2. The architecture description, ablation settings, and experimental configuration are insufficiently clear to determine where the gains come from. The unfolding backbone, dynamic transform, and inverse transform in Figure 2 closely follow Figure 4 of DISTA-Net, but inherited components are not clearly distinguished from new or modified components. Table 2 labels a variant “w/o DQM,” while the text calls it “w/o CP-DTG.” It does not specify whether NumCNN and CGPP remain in this variant or whether DISTA-Net uses identical post-processing. Thus, the ablation does not clearly establish whether the gains come from count conditioning of the thresholds. Equation (12) also does not explain how thresholds remain nonnegative. The main text states that all experiments use α=0.5\alpha=0.5, whereas the appendix gives a default of 0.7; the MSIF stage set is also inconsistently specified. Finally, two different parameter/FLOPs records are provided: 2.201M/35.247G and 2.7M/37.986G. These issues leave the configuration underlying the reported results unclear.

Questions for authors
  1. Are there controlled results comparing the full count distribution with a discrete count or count embedding, while keeping the backbone, threshold-generation architecture, and post-processing identical? If so, how large are the differences? This would help me determine whether the gains come from the particular count representation emphasized in this paper.

  2. Please identify which components in Figure 2 are inherited from DISTA-Net and which are new or modified. In particular, what inputs, connections, and operations does CP-DTG change relative to the original DTG? Does “w/o DQM” in Table 2 refer to the same variant as “w/o CP-DTG” in the text? If so, please use consistent terminology; otherwise, please explain which components each variant removes. Are NumCNN and CGPP retained in this ablation? Does DISTA-Net use the same candidate extraction and post-processing? These details would directly affect my interpretation of the ablation results in Table 2. Which α, MSIF stage set, and parameter/FLOPs record correspond to the model reporting a CSO-mAP of 47.28?

  3. Can the threshold in Equation (12) be negative in the implementation? If not, what operation ensures nonnegativity? If negative values are allowed, some responses can be amplified; how should the operation then be interpreted as shrinkage?

Reviewer_9ED1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies single-frame closely-spaced infrared small target unmixing, where overlapping point-spread functions obscure the number, sub-pixel positions, and intensities of individual sources. CISTA-Net incorporates target-count information into an ISTA-inspired unfolding network. CP-DTG combines count information with local reconstruction features to generate stage-wise thresholds. MSIF aggregates intermediate reconstructions, while the described CGPP selects candidates using the predicted count and a minimum separation constraint. On CSIST-100K, the paper reports CSO-mAP values of 47.28, 70.25, and 73.06 at c=3,5,7, improving over DISTA-Net by 0.54, 0.67, and 0.22 percentage points. At c=3, the baseline, fusion-only, CP-DTG-only, and complete variants obtain 46.74, 46.84, 47.01, and 47.28. Improvements occur at selected localization tolerances rather than uniformly improving precise localization.

Critical strengths

1.The design is motivated by the problem structure. Severe PSF overlap makes it difficult to preserve weak sources while suppressing spurious responses. Letting estimated cardinality influence this decision inside unfolding is a plausible research direction. The paper appropriately acknowledges that estimating count from the same observation adds no independent measurements and does not resolve non-identifiability by itself. 2.The reported results offer preliminary evidence of complementary components. Overall CSO-mAP improves over DISTA-Net at all three grid resolutions, and the c=3 ablations report gains from both CP-DTG and MSIF. The manuscript also acknowledges limited improvements at strict localization tolerances. These results would support the approach if their implementation and evaluation correspondence were established.

Critical weaknesses

1.The described method and supplied implementation do not currently define a consistent, verifiable pipeline. Equation (14), on page 5, specifies a bounded correction s_hat = s_N + beta50tanh(R(f)). In net/model/cistau.py, lines 364–365 and 515–519 instead imply s_final = (1+p)s_N + 0.01p*fused, with p=0.7 in the c=3 configuration. This also scales the terminal reconstruction, potentially changing candidate extraction at the fixed threshold of 50. Moreover, NumCNN logits are passed directly into INI rather than the normalized probability distribution specified in Section 4.1. The actual count-based inference path is unclear: the main c=3 evaluator leaves count filtering and NMS disabled, while count-enabled branches in cso_metric.py and detect.py read output[“count”]. In CISTAU, this field comes from annotation count_1; predictions are separately returned as x_num. An isolated synthetic execution confirms the annotation dependence of those branches. This does not establish that the reported tables used ground-truth counts, because their exact generating configurations and logs are unavailable. Nevertheless, the authors must identify the actual path and establish that ordinary inference uses predicted counts only. The supplied configurations also use 6/5/4 stages and pretrained frozen backbones, and the archive lacks the matrices, checkpoints, and complete data-loading package required to reproduce the pipeline. 2.The incremental contribution and stability of the gains are insufficiently established. The paper identifies DISTA-Net++ as already using count-guided unfolding and count-constrained inference, but provides neither a quantitative comparison nor a controlled feature-level versus threshold-level conditioning experiment. The improvements of 0.54/0.67/0.22 points are reported without repeated-seed variability or confidence intervals. Existing ablations do not separate count semantics from extra supervision/capacity, candidate filtering, or output scaling. Count accuracy and performance when counts are incorrect are also missing. Finally, c is defined as grid subdivision, yet c=3/5/7 are interpreted as physical density regimes without documenting changed target distributions. For fixed scenes, increasing c changes quantization error rather than target spacing or PSF overlap. The present evidence therefore does not sufficiently isolate the distinctive mechanism or support the density-dependent explanation.

Questions for authors

1.Which exact pipeline generated Tables 1–3? Please provide the configuration, checkpoint identifier, and training/inference logs for each relevant row. Reconcile Equation (14) with the implemented scaling, probability versus logit conditioning, stage counts, fusion weights, pretrained/frozen stages, and the consistency-loss input. Which version was evaluated? 2.What count information was used in evaluation? Please specify candidate thresholds, minimum separation, and count/NMS switches. Explain the disabled CGPP in the main configuration and the use of annotation-derived count in other metric branches. Please report matched-protocol results with no count post-processing, predicted-count post-processing, and a separately labeled oracle-count upper bound, confirming that ordinary test inference does not access annotation counts. 3.Does threshold-level count injection provide a stable advantage? Please compare against DISTA-Net++ or a controlled feature-level count-conditioning variant with the same backbone, estimator, supervision, capacity, and post-processing. Report repeated-seed results, uncertainty for the differences, count accuracy, and detection results for correct versus incorrect counts. A simple output-rescaling control would help isolate MSIF’s contribution.

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

Related papers

Discussion (0)

Sign in to comment.