Target Driven Adaptive Loss For Infrared Small Target Detection

Atsushi Ito; Takahiro Toizumi; Yuho Shoji

REVIEW 3 major objections 5 minor 29 references

Target Driven Adaptive Loss For Infrared Small Target Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A target-driven adaptive loss that weights each target by scale and local contrast is claimed to improve infrared small target detection on four public benchmarks.

desk verdict A genuinely new per-target adaptive loss for IRSTD with consistent gains, but the headline result is weakened by tuning w_T on the test split and single-run reporting. read the letter →

arxiv 2506.01349 v1 pith:GQ5YBVPQ submitted 2025-06-02 cs.CV

Yuho Shoji , Takahiro Toizumi , Atsushi Ito This is my paper

classification cs.CV

keywords infraredsmalltargetdetectiondeeplearninglossfunctionscalelocalcontrastadaptivepatch-basedmechanismIRSTD

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a new loss function, the target driven adaptive (TDA) loss, corrects two weaknesses in standard segmentation losses for infrared small target detection: tiny targets contribute almost nothing to pixel-level and image-level losses, and the hardest targets (small scale, low local contrast) are learned last. TDA loss isolates each target in its own 48×48 patch, giving every target equal weight, then increases the loss for targets whose scale or contrast falls below the training-set average. Combined with the scale-and-location-sensitive (SLS) loss at weight 0.2, it reports the best IoU and probability of detection on IRSTD-1k, SIRST-v1, SIRST-v2, and NUDT-SIRST, with the largest gains on small-scale and low-contrast targets. If correct, the loss improves detection without changing network architecture, which matters for applications such as maritime rescue and traffic monitoring.

What carries the argument

The central mechanism is the adaptive exponent $p_t$ inside a soft-IoU-based loss, computed per target patch. It combines two sigmoid terms, one for scale $s_t/s_{\mathrm{mean}}$ and one for local contrast $c_t/c_{\mathrm{mean}}$, so smaller or dimmer targets receive a larger loss weight. The patch extraction pipeline—spaghetti labeling, dilated bounding boxes, resizing to 48×48—is what makes the loss target driven: each target is treated independently and equally before the adaptive weight is applied. This is what lets multi-target scenes and hard targets contribute to training.

What would settle it

Train the same detector on IRSTD-1k with $p_t$ computed from a random permutation of the per-target scale and contrast values; if detection metrics do not fall back toward the fixed-exponent baselines, the reported gains are not caused by the scale/contrast ordering.

Watch

Extended reading notes

Core claim

The central claim is that a new loss, the target driven adaptive (TDA) loss, $L_t = -(1 - I_t^{p_t})\log(I_t)$, with per-target soft IoU $I_t$ and adaptive exponent $p_t = 1 + \sigma(-s_t/s_{\mathrm{mean}}) + \sigma(-c_t/c_{\mathrm{mean}})$, makes a detector pay attention to local regions around each target and to targets that are small or low-contrast. Each target is separated by spaghetti labeling, its bounding box dilated by a random 2–5 pixels, cropped, resized to 48×48, and the loss is averaged over all targets in the image. The exponent stays near 1 for large, high-contrast targets and grows toward 3 for small or low-contrast targets, increasing their gradient contribution. Adding this loss to the SLS loss as $L_{\mathrm{total}} = L_S + 0.2L_T$ produces the best IoU and $P_d$ on IRSTD-1k, SIRST-v1, SIRST-v2, and NUDT-SIRST, with the largest gains on small-scale and low-contrast targets.

Load-bearing premise

The loss assumes that a target's pixel count and its local contrast relative to training-set averages reliably rank how hard that target is; if those statistics are noisy or dataset-specific, the adaptive weighting would push the model toward the wrong targets.

Editorial extensions

If this is right

If the claim holds, any IRSTD segmentation network can be retrained with TDA loss added to its existing loss and gain IoU and $P_d$ without architectural changes.
Small-scale targets (under 20 pixels) and low-contrast targets, the failure cases of existing losses, become the most heavily weighted training examples.
The loss is compatible with BCE, IoU, and Dice losses in addition to SLS, so it can be plugged into a range of training setups.
The reported per-scale and per-contrast plots imply the gain is concentrated exactly where conventional losses degrade, not spread uniformly.
On the four benchmark suites, the SLS+TDA combination moves $P_d$ to 100% on SIRST-v1 and raises IoU by several points on NUDT-SIRST.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because $p_t$ is calibrated with $s_{\mathrm{mean}}$ and $c_{\mathrm{mean}}$ from the training set, the loss may need recomputation of these statistics when deployment data has a different target-size or contrast distribution; the paper does not test this transfer.
The sigmoid form is one of many possible monotone maps from scale and contrast to weight; an ablation replacing it with a simple rank-based weight would reveal whether the specific function matters or just the ordering.
The same patch-based adaptive weighting could apply to other small-object segmentation tasks, such as aerial imagery or medical imaging, but that transfer is not demonstrated here.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The thing to know: this paper proposes TDA loss, a per-target patch-based soft IoU with an exponent adapted to each target's scale and local contrast, combined with SLS loss. It reports consistent IoU and Pd gains over strong baselines on four datasets and four backbones. The per-target formulation is genuinely new, not just a tweak of known losses. The adaptive exponent is simple and plausibly does what it claims: Fig. 3 shows the largest gains on small and low-contrast targets, which is expected but still useful.

What it does well: the loss is clearly specified, the ablation of fixed vs adaptive p_t in Table 1 supports the adaptive design, and supplementary Tables 3-4 show gains across U-Net, UIU-Net, SCTransNet and with BCE/IoU/Dice, which is real evidence of generality. The comparison against six baselines on four benchmarks is solid in scope.

Where it's soft: the stress-test concern is real. Section 3 says w_T was chosen by testing 0.1-0.4 on IRSTD-1k and picking 0.2 for best IoU. That same IRSTD-1k split is then the headline row in Table 2. So the 69.50 IoU / 95.18 Pd row is not an honest out-of-sample number; part of the gain could be a selection artifact. They then reuse w_T=0.2 on the other datasets, which propagates any bias. This is the most load-bearing flaw. Second, all results are single-run point estimates with no error bars or seeds; differences of 1-2 IoU points could easily overlap. Third, the ROC curve excludes BCE and Focal with a one-line justification; that's acceptable, but it limits the ROC comparison. Also minor: the abstract says three datasets while the experiments report four. None of these are fatal to the core idea, but they need fixing.

Who it's for: researchers working on IRSTD training losses, and anyone doing segmentation with extreme class imbalance and small objects. They get a clean, testable loss formulation and a useful baseline. It deserves a serious referee, but the revision should require held-out validation or per-dataset tuning, multi-seed results, and sensitivity analysis on p_t and patch size.

Recommendation: send it to review, with the understanding that the evaluation methodology needs to be tightened before acceptance. I'd want to see the code and multi-run numbers before fully trusting the IRSTD-1k row.

Referee Report

3 major / 5 minor

Summary. The paper proposes a target-driven adaptive (TDA) loss for infrared small target detection. The loss is computed on per-target dilated patches, using a soft IoU term with an adaptive exponent p_t = 1 + sigma(-s_t/s_mean) + sigma(-c_t/c_mean) that grows for small-scale, low-contrast targets. The total loss is L_S + w_T L_T, where L_S is the SLS loss. Experiments on four datasets (IRSTD-1k, SIRST-v1, SIRST-v2, NUDT-SIRST) with MSHNet, and on three additional architectures in the supplementary material, report improved IoU and Pd over BCE, Focal, Tversky, IoU, Dice, and SLS losses. The main empirical claim is that the adaptive weighting improves detection for challenging targets while preserving overall performance.

Significance. If the result holds, the contribution is practically useful: a simple, architecture-agnostic loss that can be added to existing segmentation frameworks at no inference cost, with consistent improvements across multiple datasets and architectures in the reported tables. The paper also includes a useful mechanism check (fixed vs. adaptive p_t; stratified Pd by scale and contrast). However, the evaluation protocol currently does not exclude selection-on-test bias for the hyperparameter w_T, and all results are single-run point estimates, so the strength of the empirical claim is not yet established at the level claimed in the abstract.

major comments (3)

[Section 3, Eq. (5)] The weight w_T is selected on the IRSTD-1k dataset: the text states 'We examined w_T values between 0.1 and 0.4 using the IRSTD-1k dataset, and found that 0.2 produced the highest IoU. Therefore, we adopted this value.' The same IRSTD-1k test split is then used to report the headline IoU/Pd numbers in Table 2. As written, this does not describe a held-out validation procedure; if the selection was made on the test split, the IRSTD-1k row is not an honest out-of-sample comparison, and the gain over SLS loss (69.50 vs 67.81 IoU) may be a selection artifact. Please either tune w_T on a separate validation split and report that protocol, or report the IoU/Pd results for all tried w_T values on all datasets.
[Section 4.1, Table 2 and Table 1] All experiments appear to be single runs, with no error bars, confidence intervals, or significance tests. Many reported differences are small: for example, SIRST-v2 IoU 70.26 vs 68.88 and NUDT-SIRST Pd 98.41 vs 97.88; in Table 1, adaptive p_t beats p_t=2 by 0.27 IoU and 1.37 Pd. Without multi-seed variability, the claim that TDA loss 'achieves better detection performance' is not statistically supported. Please report means and standard deviations over at least three seeds, or a paired significance test over images.
[Section 3, Eq. (4) and Section 4.2] The adaptive mechanism rests on the ordering induced by s_t/s_mean and c_t/c_mean, but the paper gives no sensitivity analysis for the sigmoid combination or for the stability of the dataset statistics s_mean and c_mean, and the local contrast definition used in Eq. (4) ('background within the patch') differs from the definition used in Fig. 3 (background within a bounding box dilated by three pixels). Please clarify the exact c_t computation used during training and add an ablation or discussion showing that the result is not sensitive to the specific anchor statistics and functional form. This is load-bearing because Table 1 only compares one adaptive recipe to three fixed exponents; it does not test whether the adaptive ordering itself is necessary.

minor comments (5)

[Abstract] The abstract says the method is evaluated on three datasets, but the experiments in Section 4.1 use four datasets (IRSTD-1k, SIRST v1, SIRST v2, and NUDT-SIRST); please align the wording.
[Fig. 4 caption] The caption contains a typo: 'smalle scale' should be 'small scale'.
[Table 4] The table header contains 'DIce loss' where 'Dice loss' is intended.
[References] Reference [22] says 'Directed a cyclic graphs' and should be 'Directed acyclic graph'; reference [13] abbreviates the journal as 'IEEE TRGS', which should likely be 'IEEE TGRS'.
[Section 4.2, Fig. 3] The stratified Pd analysis is consistent with the loss's design, but it is a mechanism check rather than an independent confirmation of the method's overall value; the paper could state this explicitly.

Circularity Check

1 steps flagged · score 4.0 of 10

IRSTD-1k headline result is partly in-sample because w_T was selected on the same dataset; other datasets provide independent support.

fitted input called prediction [Section 3 (Eq. 5) and Table 2 (IRSTD-1k row)]
"We examined w_T values between 0.1 and 0.4 using the IRSTD-1k dataset, and found that 0.2 produced the highest IoU. Therefore, we adopted this value for w_T."

The total loss in Eq. (5) contains the weight w_T. The authors select w_T by maximizing IoU on the IRSTD-1k dataset, and then the same IRSTD-1k split is used to report the headline SLS+TDA IoU of 69.50 in Table 2. Because the paper describes no validation split, the reported IRSTD-1k result is at best ambiguous and at worst an in-sample selection artifact: the comparison against SLS loss (67.81 IoU) is not an independent out-of-sample test for that row. The result is statistically forced by the selection of w_T to maximize the reported metric.

full rationale

The loss definition itself is not circular: Eq. (2)-(4) define an adaptive weighting of per-target IoU by scale and contrast, and the improvement on small/low-contrast targets in Fig. 3 is an empirical mechanism check, not a logical consequence. No load-bearing self-citations appear; MSHNet, SLS loss, and spaghetti labeling are external. The one circular element is the selection of w_T on IRSTD-1k followed by reporting IRSTD-1k as a benchmark result. However, the central claim does not rest on that row alone: SIRST-v1, SIRST-v2, and NUDT-SIRST use the same w_T without per-dataset tuning and show consistent IoU gains, and the supplementary architecture experiments also show gains. Therefore the paper is only partially circular, not wholly reducible to its inputs.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or architectural entities. The free parameters are loss-engineering choices: the combination weight w_T, the patch size, the dilation range, and the exponent offset. The axioms are standard machine learning and domain assumptions that are not independently verified in the paper.

free parameters (4)

loss weight w_T = 0.2
Tuned on IRSTD-1k as the best IoU among 0.1 to 0.4 (Sec. 3, Eq. 5), then applied to other datasets.
patch size = 48x48
Fixed crop size for per-target patches (Sec. 3). No sensitivity analysis is provided.
dilation margin d = random integer in [2,5]
Random dilation of target bounding boxes for patch cropping (Sec. 3); acts as data augmentation, with the range chosen by hand.
p_t base offset = 1
Constant 1 in Eq. 4 sets the minimum exponent; the two sigmoid terms are added on top. Chosen by design, no ablation is shown.

assumptions (4)

domain assumption The four benchmark datasets and their standard train/test splits are representative of IRSTD, and the evaluation metrics (IoU, Pd, Fa at threshold 0.5) are standard.
Used in Sec. 4.1; the paper does not analyze dataset-specific effects or protocol sensitivity.
standard math The soft IoU defined on cropped patches is differentiable and provides useful gradients for training the segmentation model.
Used in Eq. 2; standard property of IoU-based losses.
domain assumption Deep networks learn easy patterns first and postpone harder ones, motivating the upweighting of hard targets.
Cited in Sec. 1 from Arpit et al.; this premise underpins the adaptive weighting strategy.
domain assumption Spaghetti labeling cleanly separates individual targets in the ground truth so that per-target patches can be extracted.
Used in Sec. 3 to compute per-target bounding boxes before dilation and cropping.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Target Driven Adaptive Loss For Infrared Small Target Detection." pith.science (2026). https://pith.science/paper/GQ5YBVPQ

@misc{pith2026250601349,
  author       = {Pith},
  title        = {Pith review of: Target Driven Adaptive Loss For Infrared Small Target Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GQ5YBVPQ}},
  note         = {Machine review of arXiv:2506.01349}
}

read the original abstract

We propose a target driven adaptive (TDA) loss to enhance the performance of infrared small target detection (IRSTD). Prior works have used loss functions, such as binary cross-entropy loss and IoU loss, to train segmentation models for IRSTD. Minimizing these loss functions guides models to extract pixel-level features or global image context. However, they have two issues: improving detection performance for local regions around the targets and enhancing robustness to small scale and low local contrast. To address these issues, the proposed TDA loss introduces a patch-based mechanism, and an adaptive adjustment strategy to scale and local contrast. The proposed TDA loss leads the model to focus on local regions around the targets and pay particular attention to targets with smaller scales and lower local contrast. We evaluate the proposed method on three datasets for IRSTD. The results demonstrate that the proposed TDA loss achieves better detection performance than existing losses on these datasets.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages

[1]

Target Driven Adaptive Loss For Infrared Small Target Detection

INTRODUCTION Small target detection is crucial in various applications, in- cluding maritime rescue and traffic management [1, 2]. Es- pecially, infrared small target detection (IRSTD) has gained attention because of its robustness against varying light con- ditions. IRSTD focuses on identifying targets typically oc- cupying less than 50 pixels in an imag...

work page Pith review arXiv 2025
[2]

Some prior works have introduced loss functions, such as the adversarial loss, the edge loss, and the object- detection-based loss [8, 12, 16]

RELA TED WORK Loss functions are crucial for improving detection perfor- mance. Some prior works have introduced loss functions, such as the adversarial loss, the edge loss, and the object- detection-based loss [8, 12, 16]. These losses are developed for specific network architectures. Applying these losses to different network designs is challenging, whi...

work page
[3]

The proposed TDA loss consists of two key com- ponents: a patch-based mechanism and an adaptive adjusting strategy to the scales and local contrasts

PROPOSED METHOD We design a target driven adaptive (TDA) loss to enhance de- tection performance in local regions around targets and de- tection rates for targets with smaller scales and lower local contrasts. The proposed TDA loss consists of two key com- ponents: a patch-based mechanism and an adaptive adjusting strategy to the scales and local contrast...

work page
[4]

First, we analyzed the parameterp t of the proposed TDA loss by comparing three fixed values and an adaptive set- ting

EXPERIMENTS We evaluated the proposed TDA loss through four experi- ments. First, we analyzed the parameterp t of the proposed TDA loss by comparing three fixed values and an adaptive set- ting. Next, we evaluated the detection performance of TDA loss against existing loss functions on four datasets. We also evaluated the robustness across different scale...

work page arXiv
[5]

The proposed TDA loss comprises two key compo- nents: a patch-based feature learning mechanism, and a dy- namic loss adaptation strategy to individual target character- istics

CONCLUSION In this study, we introduced the TDA loss for infrared small target detection, designed to enhance detection performance, particularly for challenging small scale and low local contrast targets. The proposed TDA loss comprises two key compo- nents: a patch-based feature learning mechanism, and a dy- namic loss adaptation strategy to individual ...

work page
[6]

Single-frame infrared small-target detection: A survey,

Mingjing Zhao, Wei Li, Lu Li, Jin Hu, Pengge Ma, and Ran Tao, “Single-frame infrared small-target detection: A survey,”IEEE TGRS, vol. 10, no. 2, pp. 87–119, 2022

work page 2022
[7]

Review of dim small target detection algorithms in single-frame infrared images,

Ke Zhang, Shuyan Ni, Dashuang Yan, and Aidi Zhang, “Review of dim small target detection algorithms in single-frame infrared images,” inIEEE IMCEC. IEEE, 2021, vol. 4, pp. 2115–2120

work page 2021
[8]

Pay attention to local contrast learning networks for infrared small target detection,

Chuang Yu, Yunpeng Liu, Shuhang Wu, Xin Xia, Zhuhua Hu, Deyan Lan, and Xin Liu, “Pay attention to local contrast learning networks for infrared small target detection,”IEEE GRSL, vol. 19, pp. 1–5, 2022

work page 2022

Show all 29 references

[9]

Attentional local contrast networks for infrared small target detection,

Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard, “Attentional local contrast networks for infrared small target detection,”IEEE TGRS, vol. 59, no. 11, pp. 9813– 9824, 2021

2021
[10]

RISTDnet: Robust infrared small target detection network,

Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Haoliang Zheng, and Wei Zhang, “RISTDnet: Robust infrared small target detection network,”IEEE GRSL, vol. 19, pp. 1–5, 2021

2021
[11]

A local contrast method for infrared small-target detection utilizing a tri- layer window,

Jinhui Han, Saed Moradi, Iman Faramarzi, Chengyin Liu, Honghui Zhang, and Qian Zhao, “A local contrast method for infrared small-target detection utilizing a tri- layer window,”IEEE GRSL, vol. 17, no. 10, pp. 1822– 1826, 2019

2019
[12]

Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection,

Yimian Dai and Yiquan Wu, “Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection,”IEEE J-STARS, vol. 10, no. 8, pp. 3752–3767, 2017

2017
[13]

ISNet: Shape matters for in- frared small target detection,

Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo, “ISNet: Shape matters for in- frared small target detection,” inCVPR, 2022, pp. 877– 886

2022
[14]

Infrared small target detection with scale and location sensitivity,

Qiankun Liu, Rui Liu, Bolun Zheng, Hongkui Wang, and Ying Fu, “Infrared small target detection with scale and location sensitivity,” inCVPR, 2024, pp. 17490– 17499

2024
[15]

Dense nested attention network for infrared small tar- get detection,

Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo, “Dense nested attention network for infrared small tar- get detection,”IEEE TIP, vol. 32, pp. 1745–1758, 2022

2022
[16]

UIU- Net: U-Net in U-Net for infrared small object detec- tion,

Xin Wu, Danfeng Hong, and Jocelyn Chanussot, “UIU- Net: U-Net in U-Net for infrared small object detec- tion,”IEEE TIP, vol. 32, pp. 364–376, 2022

2022
[17]

Miss detec- tion vs. false alarm: Adversarial learning for small ob- ject segmentation in infrared images,

Huan Wang, Luping Zhou, and Lei Wang, “Miss detec- tion vs. false alarm: Adversarial learning for small ob- ject segmentation in infrared images,” inICCV, 2019, pp. 8509–8518

2019
[18]

SCTransNet: Spatial-channel cross transformer network for infrared small target detection,

Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian, “SCTransNet: Spatial-channel cross transformer network for infrared small target detection,” IEEE TRGS, 2024

2024
[19]

Asymmetric contextual modulation for infrared small target detection,

Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard, “Asymmetric contextual modulation for infrared small target detection,” inCVPRW, 2021, pp. 950–959

2021
[20]

A closer look at memorization in deep networks,

Devansh Arpit et al., “A closer look at memorization in deep networks,” inICML, 2017, pp. 233–242

2017
[21]

Interior Attention-Aware Network for Infrared Small Target Detection,

Kewei Wang, Shuaiyuan Du, Chengxin Liu, and Zhiguo Cao, “Interior Attention-Aware Network for Infrared Small Target Detection,”IEEE TGRS, vol. 60, pp. 1–13, 2022

2022
[22]

One-stage cascade refinement networks for infrared small target detection,

Yimian Dai, Xiang Li, Fei Zhou, Yulei Qian, Yaohong Chen, and Jian Yang, “One-stage cascade refinement networks for infrared small target detection,”IEEE TGRS, vol. 61, pp. 1–17, 2023

2023
[23]

Batching soft IoU for training semantic segmentation networks,

Yifeng Huang, Zhirong Tang, Dan Chen, Kaixiong Su, and Chengbin Chen, “Batching soft IoU for training semantic segmentation networks,”IEEE Signal Process. Lett., vol. 27, pp. 66–70, 2019

2019
[24]

Generalised dice over- lap as a deep learning loss function for highly unbal- anced segmentations,

Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso, “Generalised dice over- lap as a deep learning loss function for highly unbal- anced segmentations,” inDLMIA-MLCDS. Springer, 2017, pp. 240–248

2017
[25]

Focal Loss for dense object detec- tion,

Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar, “Focal Loss for dense object detec- tion,”IEEE TPAMI, vol. 42, no. 2, pp. 318–327, 2020

2020
[26]

Rethinking in- tersection over union for small object detection in few- shot regime,

Pierre Le Jeune and Anissa Mokraoui, “Rethinking in- tersection over union for small object detection in few- shot regime,”arXiv preprint arXiv:2307.09562, 2023

2023 arXiv
[27]

Spaghetti labeling: Directed a cyclic graphs for block-based connected components labeling,

Federico Bolelli, Stefano Allegretti, Lorenzo Baraldi, and Costantino Grana, “Spaghetti labeling: Directed a cyclic graphs for block-based connected components labeling,”IEEE TIP, vol. 29, pp. 1999–2012, 2020

1999
[28]

Tversky loss function for image seg- mentation using 3D fully convolutional deep networks,

Seyed Sadegh Mohseni Salehi, Deniz Erdogmus, and Ali Gholipour, “Tversky loss function for image seg- mentation using 3D fully convolutional deep networks,” inMLMI. Springer, 2017, pp. 379–387

2017
[29]

U-Net: Convolutional networks for biomedical image segmentation,

Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-Net: Convolutional networks for biomedical image segmentation,” inMICCAI. Springer, 2015, pp. 234– 241. TARGET DRIVEN ADAPTIVE LOSS FOR INFRARED SMALL TARGET DETECTION Supplementary Material Table 3: We compared our propos...

2015

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Target Driven Adaptive Loss For Infrared Small Target Detection

INTRODUCTION Small target detection is crucial in various applications, in- cluding maritime rescue and traffic management [1, 2]. Es- pecially, infrared small target detection (IRSTD) has gained attention because of its robustness against varying light con- ditions. IRSTD focuses on identifying targets typically oc- cupying less than 50 pixels in an imag...

work page Pith review arXiv 2025

[2] [2]

Some prior works have introduced loss functions, such as the adversarial loss, the edge loss, and the object- detection-based loss [8, 12, 16]

RELA TED WORK Loss functions are crucial for improving detection perfor- mance. Some prior works have introduced loss functions, such as the adversarial loss, the edge loss, and the object- detection-based loss [8, 12, 16]. These losses are developed for specific network architectures. Applying these losses to different network designs is challenging, whi...

work page

[3] [3]

The proposed TDA loss consists of two key com- ponents: a patch-based mechanism and an adaptive adjusting strategy to the scales and local contrasts

PROPOSED METHOD We design a target driven adaptive (TDA) loss to enhance de- tection performance in local regions around targets and de- tection rates for targets with smaller scales and lower local contrasts. The proposed TDA loss consists of two key com- ponents: a patch-based mechanism and an adaptive adjusting strategy to the scales and local contrast...

work page

[4] [4]

First, we analyzed the parameterp t of the proposed TDA loss by comparing three fixed values and an adaptive set- ting

EXPERIMENTS We evaluated the proposed TDA loss through four experi- ments. First, we analyzed the parameterp t of the proposed TDA loss by comparing three fixed values and an adaptive set- ting. Next, we evaluated the detection performance of TDA loss against existing loss functions on four datasets. We also evaluated the robustness across different scale...

work page arXiv

[5] [5]

The proposed TDA loss comprises two key compo- nents: a patch-based feature learning mechanism, and a dy- namic loss adaptation strategy to individual target character- istics

CONCLUSION In this study, we introduced the TDA loss for infrared small target detection, designed to enhance detection performance, particularly for challenging small scale and low local contrast targets. The proposed TDA loss comprises two key compo- nents: a patch-based feature learning mechanism, and a dy- namic loss adaptation strategy to individual ...

work page

[6] [6]

Single-frame infrared small-target detection: A survey,

Mingjing Zhao, Wei Li, Lu Li, Jin Hu, Pengge Ma, and Ran Tao, “Single-frame infrared small-target detection: A survey,”IEEE TGRS, vol. 10, no. 2, pp. 87–119, 2022

work page 2022

[7] [7]

Review of dim small target detection algorithms in single-frame infrared images,

Ke Zhang, Shuyan Ni, Dashuang Yan, and Aidi Zhang, “Review of dim small target detection algorithms in single-frame infrared images,” inIEEE IMCEC. IEEE, 2021, vol. 4, pp. 2115–2120

work page 2021

[8] [8]

Pay attention to local contrast learning networks for infrared small target detection,

Chuang Yu, Yunpeng Liu, Shuhang Wu, Xin Xia, Zhuhua Hu, Deyan Lan, and Xin Liu, “Pay attention to local contrast learning networks for infrared small target detection,”IEEE GRSL, vol. 19, pp. 1–5, 2022

work page 2022