Accepted by the 19th European Conference on Computer Vision (ECCV 2026)

SPIRE

Rethinking IRSTD: Single-Point Supervision Guided Encoder-Only Framework Is Enough for Infrared Small Target Detection

重新思考红外小目标检测：单点监督引导的纯编码器框架足以实现红外小目标检测

倪日祥（Rixiang Ni）陈军（Jun Chen）李博扬（Boyang Li） et al.

College of Electronic Science and Technology

National University of Defense Technology · Wei An Research Group

TL;DR: SPIRE turns centroid points into physically grounded probabilistic targets, regresses them with a lightweight high-resolution encoder, and localizes targets by peak extraction.

Overview

Infrared small target detection is commonly framed as dense mask reconstruction, with a decoder progressively restoring spatial resolution. SPIRE revisits that choice: when the operational goal is reliable centroid localization, supervision, representation, and output can all be centered on a probabilistic response instead of a reconstructed contour.

The method follows three steps: construct a structured response target from each centroid annotation, regress the response map on a fixed high-resolution branch, and recover target coordinates through local-maximum detection and sub-pixel refinement. The paper figure above opens at full resolution.

01 Single-Point Supervision

Construct a compact, center-enhanced regression target from each annotated centroid.

02 Probabilistic Response Encoding

Regress the response on a fixed high-resolution branch without a decoder or skip connections.

03 Peak-Based Localization

Read target centroids directly through local maxima and sub-pixel refinement.

How PRPS Constructs the Regression Target

PRPS uses the normalized local radiometric profile C_k to modulate an unchanged Gaussian point-response prior G_k. Panel I uses an enlarged three-dimensional UAV-like surface to make this weighting visually explicit; during training, C_k is computed from the original infrared image patch.

I Illustrative UAV Radiometric Profile

A magnified three-dimensional surface of local target radiometry—not a mask or measured network response.

C_k

Three-dimensional unit-peak isotropic Gaussian point-response prior — II **Gaussian Point-Response Prior**
The unchanged unit-peak isotropic Gaussian is centered at the refined radiometric peak.
`G_k`

Three-dimensional contrast-aware PRPS target after normalized element-wise modulation — III **Gaussian-Weighted PRPS Target**
The weighted response retains local radiometric asymmetry while concentrating near the center.
`H_k = Norm(G_k ⊙ C_k)`

C_k × G_k → H_k

Schematic visualization of the paper-defined construction. In training, C_k comes from the min-max normalized local image patch; σ=2 and r=6 follow the paper. No target contour is reconstructed.

View figure provenance

Qualitative Results Visualization

Explore verified, precomputed SPIRE outputs on three selected samples from the official SIRST-UAVB test split. All results use the released checkpoint and the same fixed evaluation settings.

Verified precomputed results

SIRST-UAVB sample 2434 input infrared image

SPIRE centroid prediction on SIRST-UAVB sample 2434

Matched Loading verified localization record

View run manifest

Fixed settings: 640 × 640 · threshold 0.35 · δ = 5 px

Unseen real scenes

Real-World Video Demonstrations

Both videos use the SPIRE checkpoint trained on the SIRST-UAVB (2400:600) train/test split. Neither real-scene sequence was included in training. The successful detections provide qualitative evidence of cross-scene generalization beyond the benchmark; these clips are not part of the benchmark evaluation.

Real Scene I · Structured Background Unseen during training

A small target is localized amid dense texture and a strong linear background feature.

Real Scene II · Mountainous Terrain Unseen during training

A bright airborne target is localized against a high-contrast terrain background.

Red marker: SPIRE-predicted target location

Reported Results

Tables 1 and 2 reproduce all cross-method comparisons reported in the final paper. Table 1 follows the unified centroid protocol at δ = 5 pixels. Table 2 evaluates matching-threshold robustness on SIRST4 and reports throughput under the same input and hardware setting. In Table 1, bold and underlined values mark the best and second-best entries computed from the reported numbers.

Table 1 · Main benchmark

Comparison with representative methods

Pre, Rec (Pd), and F1 are reported in %. Fa is reported in units of 10⁻⁸. FLOPs and Params are measured at 640 × 640 input resolution.

Comparison with representative methods on SIRST-UAVB and SIRST4 under the unified centroid protocol.
Method	Venue	SIRST-UAVB (2400:600)				SIRST4 (2285:1067)				FLOPs (G) ↓	Params (M) ↓
Method	Venue	Pre ↑	Rec (Pd) ↑	F1 ↑	Fa ↓	Pre ↑	Rec (Pd) ↑	F1 ↑	Fa ↓	FLOPs (G) ↓	Params (M) ↓
ACM [11]	WACV '21	87.01	71.16	78.29	34.04	90.17	70.38	79.05	44.08	2.51	0.40
ALCNet [12]	TGRS '21	95.06	81.11	87.53	12.72	91.45	72.41	80.82	38.90	2.36	0.43
ISTDU-Net [16]	GRSL '22	79.28	78.08	78.67	61.54	93.28	72.03	81.29	29.83	49.65	2.75
RDIAN [24]	TGRS '23	52.68	77.91	62.86	211.08	90.64	80.83	85.45	47.98	23.24	0.22
DNANet [17]	TIP '23	94.48	89.54	91.95	15.77	93.99	81.20	87.13	29.82	89.13	4.70
SCTransNet [37]	TGRS '24	98.27	95.95	97.09	5.09	81.20	86.39	83.72	114.98	63.22	11.19
MSHNet [21]	CVPR '24	71.90	88.02	79.09	104.27	90.68	92.86	91.47	60.51	38.16	4.07
SDSNet [38]	TGRS '25	97.60	96.29	96.94	7.12	87.35	88.27	87.81	73.48	42.42	2.49
L²SKNet [28]	TGRS '25	95.83	89.04	92.31	11.70	92.89	91.42	92.16	40.20	43.09	0.90
SPIRE (Ours)	—	99.82	94.44	97.05	1.02	95.00	94.21	94.60	28.53	7.68	0.29

Table 2 · Robustness and speed

Multi-threshold comparison on SIRST4

Rec (Pd) and F1 are reported in %. Fa is reported in units of 10⁻⁸. Results at δ = 5, 8, and 10 are identical; FPS follows the paper's shared input and hardware setting.

Multi-threshold robustness and inference speed comparison on SIRST4.
Method	δ = 3 px			δ = 5, 8, 10 px			FPS ↑
Method	Rec (Pd) ↑	F1 ↑	Fa ↓	Rec (Pd) ↑	F1 ↑	Fa ↓	FPS ↑
DNANet [17]	80.52	86.41	33.72	81.20	87.13	29.82	80.86
SCTransNet [37]	80.15	81.18	99.42	86.39	83.72	114.98	67.09
SDSNet [38]	86.39	87.21	67.43	88.27	87.81	73.48	46.97
SPIRE (Ours)	92.93	94.74	18.61	94.21	94.60	28.53	261.2

0.29MParameters

7.68GFLOPs · 640²

261.2FPS · paper setting

All values are transcribed from the final paper. FLOPs and Params use 640 × 640 input resolution; Table 2 throughput is paper-reported and is not re-measured in the browser.

Citation

If this work supports your research, please consider citing the paper. See the Paper and Code.

@inproceedings{ni2026spire,
  title={Rethinking IRSTD: Single-Point Supervision Guided
         Encoder-Only Framework Is Enough for Infrared
         Small Target Detection},
  author={Ni, Rixiang and Chen, Jun and Li, Boyang and
          Li, Yonghao and He, Wujiao and Wang, Yuji and
          Ren, Feiyu and Yuan, Haoyang and An, Wei},
  booktitle={European Conference on Computer Vision},
  year={2026}
}