acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Lightweight Infrared Small Moving Target Detection under Heavy Noise with Temporal Templates and Neighborhood Evidence

Abstract

Infrared small moving targets occupy only a few pixels, lack texture and shape information, and are often buried in strong noise and complex backgrounds, making reliable detection highly challenging. Existing multi-frame methods improve detection by exploiting inter-frame motion, and per-pixel temporal methods among them achieve leading results. However, these methods generally process multi-frame features throughout the network, resulting in slow inference and high memory consumption, and they struggle to distinguish targets from noise fluctuations under heavy noise, where the temporal change of a target pixel is often weaker than that of a noise pixel. We observe that a moving target changes neighboring pixels one after another, whereas a noise fluctuation stays on a single pixel, and propose TRACE, a lightweight detection network built on this difference. After standardizing each pixel against its own intensity history in the temporal window, TRACE compresses the multi-frame sequence into two-dimensional features using shared learnable temporal templates in the first layer, accumulating weak target signals while avoiding layer-by-layer multi-frame processing. It then aggregates evidence over a small spatial neighborhood, following the target track and suppressing false alarms caused by noise. Experiments on public datasets show that TRACE outperforms representative multi-frame methods, particularly under heavy noise, and that, compared with the leading per-pixel temporal method, it requires fewer parameters and less memory while running faster. This work offers a new perspective on efficient and robust infrared small moving target detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
infrared small target detection, moving target detection, multi-frame detection, heavy noise, temporal template matching, spatial aggregation, lightweight network, real-time inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_caioOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Based on the observation that "target pixels sequentially alter the brightness of surrounding pixels, whereas noise fluctuations occur only at individual pixels," the authors developed two modules, TMM and LSA, to enable rapid detection.

Critical strengths

The detection speed is high. This is a strong starting point, as current multi-frame detection methods indeed require more lightweight and efficient solutions, and only such approaches can be more effectively deployed in practice. This constitutes a sound and promising foundation for the work.

Critical weaknesses

At first glance, this paper appears to be well organized; however, upon closer examination, its readability is poor. The abstract and introduction do not clearly articulate the research problem itself, and the content appears to be a restatement of existing methods. The methodology section is particularly difficult to follow and lacks logical coherence.

  1. The motivation presented in the introduction has already been extensively discussed in the literature, and it does not require further elaboration here. The references you cite already address this point. As such, this does not constitute the specific motivation for your study.
  2. The writing of this paper requires comprehensive refinement, as the current text is quite difficult to read. Although the author’s intent is discernible, the argument only becomes clear after reading the entire manuscript and then revisiting it from the beginning. Moreover, the paper lacks a central thesis that effectively synthesizes its main points. It is recommended that the author revise the manuscript to improve clarity and coherence.
  3. I have not been able to identify the motivation for TTM and LSA in the key sections. Why is that? The methods section appears to be a restatement of the introduction.
Questions for authors
  1. What is the motivation behind the TMM and LSA?
  2. What is the rationale behind setting the hyperparameters in the objective function in this manner?
  3. How does LSA aggregate motion trajectories? I notice that the network architecture employs a depthwise convolution, which I find difficult to understand.
  4. Spatial and temporal features are modeled separately, and temporal motion context is utilized to guide spatial feature extraction. This approach has been adopted by many authors, and the key distinction among them—such as in methods like MOCID—lies in how the temporal context is integrated and how the spatial features are subsequently refined.

There are a few minor issues, such as text overlapping with figure; it is recommended that the entire document be checked for such problems.

Reviewer_yyb7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies infrared small moving target detection under heavy noise. The method first uses PTS to standardize each pixel's temporal sequence, then uses TTM to compress information from 40 frames into 32 two-dimensional feature channels in a single step, and finally uses LSA to aggregate template responses within a neighborhood. The main idea is that a moving target leaves temporally related signals at neighboring positions, while isolated noise usually lacks this neighborhood support.

The paper reports a detection rate of 76.34% on HiNo at a false-alarm rate of 1.76 per 100,000 pixels, about 17.2 percentage points above the DeepPro version used for comparison. It also reports 711.5 FPS under the specified timing conditions. The appendices provide results from multiple training runs with different random seeds, results for each sequence, the procedure for setting detection thresholds on the training set, additional experiments addressing failed training of ablation variants, and results on RGBT-Tiny.

Critical strengths
  1. The model is small and outperforms the compared DeepPro version across multiple training runs. Removing the time dimension after the first layer reduces the cost of later processing, making the design useful for detection with limited computing resources. On HiNo, all five runs with different random seeds outperform the compared DeepPro version. Appendix E.2 also provides results for each sequence and an estimated interval for the improvement obtained by resampling, giving a clearer picture of the stability of the results. The combination of a compact model and consistent gains is a meaningful practical contribution.

  2. The appendices include additional checks on threshold selection and failed ablation training. In E.1, a threshold determined by cross-validation on the training set and then fixed, together with point detection, gives a detection rate of 79.09% and 0.92 false alarms per 100,000 pixels on HiNo. This shows that good results can also be obtained without selecting thresholds on the test set. To address failed training of the original ablation variant, E.5 adds experiments that temporarily omit the budget-related loss terms at the start of training or remove them throughout training. Once training succeeds, neighborhood aggregation still gives a clear improvement. E.8 also compares five training runs on real thermal-infrared data.

Critical weaknesses
  1. The comparison with directly related methods is incomplete. Section IV-E of the cited DeepPro paper already includes DeepPro-Plus and DeepPro+S, both of which add spatial information. The published HiNo results give DeepPro-Plus a detection rate of 76.23% and 1.69 false alarms per 100,000 pixels, and DeepPro+S 76.52% and 1.03, respectively. TRACE's main table reports 76.34% and 1.76. The paper should compare with these versions and discuss how they relate to LSA. The approximately 17.2-point gain should therefore be understood as an improvement over the evaluated DeepPro version. TRACE's compact design remains useful, but its accuracy, computation, and GPU memory trade-offs against these other versions need to be established under the same conditions.

  2. The limits of fixed thresholds and the timing results need to be clearer. The main table and the comparisons across noise conditions select thresholds using test-set false-alarm rates. This allows detection rates to be compared at the same false-alarm rate, but does not directly show that thresholds would need no adjustment in practice. The fixed-threshold point results in E.1 are encouraging, reaching 79.09% detection with 0.92 false alarms per 100,000 pixels on HiNo. However, under H4 in Table 9, point detection gives 3.84 false alarms per 100,000 pixels, above the specified limit of 1.76. In Table 4, raw-pixel detection with a fixed threshold gives 3.44 false alarms per 100,000 pixels on HiNo. These results call for a clearer explanation of when threshold adjustment is needed. In addition, Appendix F times the methods using random inputs, several comparison models have untrained weights, and the numbers of detections returned by the methods differ substantially. The reported speed ratios should be interpreted within those timing settings until they are checked on real videos.

Questions for authors
  1. Please explain why DeepPro-Plus and DeepPro+S were not compared, and add comparisons using the same data split, detection procedure, false-alarm requirement, and hardware. How does TRACE compare in detection rate, GPU memory use, and processing speed?

  2. Based on the loss ablations in Appendices E.5 and E.8, could the authors clarify how much of the improvement comes from the network structure and how much from the budget-related loss terms? Would adding the same terms to the comparison models be expected to change the comparison?

  3. Given the fixed-threshold results in Tables 4 and 9, under what noise conditions should the threshold be adjusted, and how could that adjustment be made without using test annotations?

  4. Could the authors report processing speed on real test videos using trained models and the same timing procedure, including preprocessing, model computation, and detection extraction, and separately report the delay from waiting for 20 future frames?

Reviewer_8eoBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents TRACE, a lightweight multi-frame infrared small moving target detection framework designed for challenging low-signal-to-noise-ratio scenarios. The proposed method integrates three components: Per-pixel Temporal Standardization (PTS) for normalizing temporal intensity variations, Temporal Template Matching (TTM) for extracting compact temporal representations from 40-frame sequences, and Local Spatial Aggregation (LSA) for enhancing target discrimination through spatial evidence aggregation. Experiments on NUDT-MIRSDT, NUDT-MIRSDT-HiNo, IRDST-simulation, and selected real thermal-infrared sequences from RGBT-Tiny demonstrate improved detection performance under heavy noise, alongside reduced model complexity, GPU memory usage, and inference time compared with several existing methods. However, the methodological novelty, generalizability of the underlying assumptions, and sufficiency of the experimental validation remain open concerns.

Critical strengths

The paper addresses a relevant problem in infrared small moving target detection and reports encouraging performance-efficiency trade-offs. The evaluation includes multiple datasets, noise settings, and ablation studies.

Critical weaknesses

The proposed TTM and LSA modules primarily rely on established operations, including learned temporal projections, depthwise and dilated convolutions, and residual connections. In particular, TTM closely resembles conventional temporal filtering or linear feature projection, while LSA implicitly aggregates local spatial responses without explicitly modeling target motion or trajectories. Although the proposed combination achieves promising performance and computational efficiency, the paper does not sufficiently clarify its methodological novelty over existing lightweight architectures. Moreover, the ablation studies lack comparisons with parameter-matched temporal convolutions, direct temporal projections, and alternative spatial aggregation mechanisms, making it difficult to establish the necessity and individual advantages of the proposed designs.

TRACE relies on the assumption that moving targets exhibit coherent spatiotemporal responses while noise fluctuations remain spatially localized. However, this distinction may not hold under realistic infrared imaging conditions involving correlated sensor noise, camera motion, structured backgrounds, and persistent detector artifacts. The evaluated noise models are primarily based on synthetic perturbations, while the real-data experiments provide limited evidence of robustness across different sensors and environments. Furthermore, the method's sensitivity to target velocity, nonlinear trajectories, and nearly stationary targets remains insufficiently investigated, particularly given the limited spatial aggregation range relative to the 40-frame temporal window. Additionally, the non-causal design requires 20 future frames, introducing detection latency that is not adequately reflected in the reported inference efficiency. These limitations leave the practical generalizability of the proposed framework insufficiently established.

Questions for authors

The physical realism of the noise model and the motivation of this work raise significant concerns. To the best of this reviewer's understanding, modern infrared imaging systems under typical operating conditions generally do not exhibit temporally independent random noise of the magnitude introduced in the synthetic HiNo benchmark. Although the authors attempt to simulate challenging imaging conditions by combining several noise components, the resulting degradation appears to deviate substantially from the noise characteristics commonly encountered in practical infrared small target detection. Consequently, it is questionable whether the proposed experimental setting reflects a representative real-world challenge. More fundamentally, constructing an unusually severe synthetic noise scenario, under which existing state-of-the-art methods perform poorly, and then proposing a method specifically tailored to its statistical properties is not, by itself, a convincing demonstration of progress on practical infrared small target detection.

Furthermore, the authors state that "a moving target changes neighboring pixels one after another, whereas a noise fluctuation stays on a single pixel." This assumption largely reflects the properties of independent Gaussian noise rather than the physical characteristics of practical infrared imaging systems. Infrared imaging has distinct noise characteristics compared with visible-light imaging, particularly detector nonuniformity and its residual artifacts after correction, such as stripe noise, screen-door patterns, and spatially correlated bias. We believe that infrared small target detection research should be grounded in realistic imaging physics rather than artificially intensified random noise. If the authors aim to exploit target motion continuity, evaluating genuinely weaker targets under realistic infrared noise and interference would be more meaningful. The authors should therefore justify the physical relevance of their noise model and demonstrate the effectiveness of TRACE under practical infrared imaging conditions.

The novelty of the proposed TTM and LSA modules appears limited. TTM essentially performs temporal channel projection using a standard 1×1 convolution, while LSA consists of conventional depthwise, pointwise, and dilated convolutions with residual connections. Combining existing modules can indeed create a valuable architecture, and the existing experiments seem to demonstrate the value of some components. However, compared with existing works such as DeepPro, the improvement on the original dataset is very limited. Although the synthetic noise benchmark is adopted from previous work, as we have emphasized earlier, its noise model clearly does not conform to typical infrared imaging conditions. Strong noise may occur in infrared imaging under certain circumstances, but it is often dominated by fixed-pattern components, for which many compensation and correction methods are available.

The authors emphasize enhancing the exploitation of target trajectory information. However, the proposed method appears to rely primarily on LSA to implicitly aggregate local features, without explicitly modeling or enhancing target trajectories. The current results do not convincingly demonstrate that the model achieves the claimed trajectory-aware enhancement. Instead, its effectiveness seems to stem more from exploiting local temporal variations and spatial feature aggregation than from genuine trajectory modeling.

The architectural novelty of LSA appears limited. Although the authors provide some ablation studies, they mainly compare the use of spatial neighborhood information against replacing all spatial convolutions with 1×1 convolutions. This only demonstrates the importance of local spatial information, rather than justifying the specific LSA design. Further analysis is needed to determine the appropriate receptive field size, the necessity of multiple convolutional blocks, and whether simpler alternatives, such as using dilated convolutions directly, could achieve comparable performance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.