TY-RIST: Infrared Small Target Detection

• TY-RIST is an infrared small target detection method designed for tiny, low-contrast objects, employing a stride-aware backbone and a high-resolution P2 branch.
• It improves precision and recall by integrating three cascaded Coordinate Attention blocks and replacing CIoU with Normalized Gaussian Wasserstein Distance for bounding-box regression.
• The architecture uses strategic branch pruning and PAN removal to achieve up to 123 FPS while boosting mAP@50 by 7.9% on benchmark datasets.

Searching arXiv for the exact TY-RIST paper and closely related name collisions to ground the article in current arXiv metadata. TY-RIST most specifically denotes “Tactical YOLO Tricks for Real-time Infrared Small Target Detection,” an infrared small target detection architecture built as an optimized re-design of YOLOv12n for tiny, low-contrast, low-SNR objects in cluttered infrared backgrounds (Atrash et al., 26 Sep 2025). The method combines a stride-aware backbone, a high-resolution detection branch and head built from shallow features, cascaded Coordinate Attention blocks, branch pruning, and Normalized Gaussian Wasserstein Distance for bounding-box regression. In the reported experiments, it improves $[email protected]$ IoU by $+7.9\%$, Precision by $+3\%$, and Recall by $+10.2\%$, while achieving up to $123$ FPS on a single GPU; cross-dataset validation on a fifth dataset is reported as further evidence of generalization (Atrash et al., 26 Sep 2025).

1. Problem setting and design objective

Infrared small target detection (IRSTD) is characterized in the reported formulation by objects occupying less than $0.5\%$ of image area, low contrast $(<0.15)$, and low SNR (Atrash et al., 26 Sep 2025). The central difficulties are target loss under aggressive downsampling, false alarms induced by textured or cluttered infrared backgrounds, missed detections caused by weak saliency, and the need to preserve real-time throughput on constrained platforms.

TY-RIST addresses these failure modes through five coordinated modifications to YOLOv12n: reducing the first backbone stride from $2$ to $1$, adding a shallow-feature high-resolution branch and detection head $(P2)$, inserting three cascaded Coordinate Attention blocks on that branch, pruning branches and the PAN neck when they do not help, and replacing CIoU with Normalized Gaussian Wasserstein Distance (NWD) for regression (Atrash et al., 26 Sep 2025). The overall design goal is not merely higher accuracy, but higher accuracy under the specific small-target regime where standard multi-scale detection pipelines may lose critical detail through early spatial compression.

This design can be read as a redistribution of representational capacity toward the highest-resolution pathway. That interpretation is consistent with the reported ablations, which show repeated gains when preserving or emphasizing shallow, spatially dense feature maps and when removing components that mainly serve larger-object scales (Atrash et al., 26 Sep 2025).

2. Architectural structure

TY-RIST starts from YOLOv12n and changes the first convolutional block stride from $+7.9\%$0 to $+7.9\%$1 (Atrash et al., 26 Sep 2025). This doubles the spatial resolution of downstream feature maps relative to the vanilla model without enlarging the input. The paper reports that tiny targets remain sharp at stride $+7.9\%$2 but blur at stride $+7.9\%$3, and this observation is used to justify the stride-aware backbone.

The architecture explicitly incorporates the shallow $+7.9\%$4 feature map into the neck to generate an additional pyramid level $+7.9\%$5, which is used by an added x-small detection head (Atrash et al., 26 Sep 2025). In the untrimmed topology, the network therefore contains the standard YOLOv12n heads $+7.9\%$6, $+7.9\%$7, and $+7.9\%$8, together with the added high-resolution $+7.9\%$9 head. The heads are anchor-free decoupled detection modules following YOLOv12 defaults, and for single-class IRSTD the classification dimension is $+3\%$0 (Atrash et al., 26 Sep 2025).

Three Coordinate Attention blocks are inserted serially on the $+3\%$1 branch immediately before prediction (Atrash et al., 26 Sep 2025). Their stated function is to enrich weak tiny-target features while preserving positional information. The neck initially uses YOLOv12n’s FPN+PAN, but the method allows dataset-specific trimming. For ITSDT-15k, the PAN is removed and non-$+3\%$2 heads are disabled in inference; for NUAA-SIRST, partial PAN is retained to build the feature needed by $+3\%$3, because larger objects occur more often there (Atrash et al., 26 Sep 2025).

A concise way to describe the final pipeline is: input $+3\%$4 stride-aware backbone $+3\%$5 feature maps $+3\%$6 $+3\%$7 extended neck producing $+3\%$8 in addition to the standard levels $+3\%$9 cascaded Coordinate Attention on $+10.2\%$0 $+10.2\%$1 anchor-free decoupled detection heads (Atrash et al., 26 Sep 2025).

3. Mathematical components

TY-RIST replaces CIoU with NWD in the regression term of the YOLOv12 detection loss (Atrash et al., 26 Sep 2025). Bounding boxes are modeled as $+10.2\%$2D Gaussians, and the paper uses the following Wasserstein-based similarity:

$+10.2\%$3

For bounding boxes $+10.2\%$4 and $+10.2\%$5, the simplified form adopted is

$+10.2\%$6

The normalized similarity is then

$+10.2\%$7

where $+10.2\%$8 is a dataset-dependent constant tuned by ablation (Atrash et al., 26 Sep 2025). The total detection loss preserves YOLOv12’s structure but substitutes the regression term:

$+10.2\%$9

The paper motivates this replacement by the instability of IoU-based losses on tiny boxes, where small positional deviations can produce large changes in overlap and even zero-overlap cases (Atrash et al., 26 Sep 2025).

The Coordinate Attention blocks on the $123$0 branch are defined through directional pooling and multiplicative gating. Given $123$1,

$123$2

After concatenation and bottleneck transformation,

$123$3

the attention maps are produced and applied as

$123$4

TY-RIST applies three such CA blocks sequentially on the $123$5 branch (Atrash et al., 26 Sep 2025). This serial arrangement is reported to progressively enhance weak tiny-target features and improve recall.

4. Pruning strategy, training protocol, and runtime configuration

A defining feature of TY-RIST is that it treats architectural pruning as part of the detector design rather than as a post hoc compression step (Atrash et al., 26 Sep 2025). The pruning procedure begins with empirical head ablation: individual heads are disabled during inference to measure their contribution. On ITSDT-15k, pruning $123$6 retains or slightly improves performance, and the PAN path is replaced with identity/bypass. On NUAA-SIRST, by contrast, $123$7 are retained and PAN is only partially removed because larger objects still benefit from an additional scale (Atrash et al., 26 Sep 2025).

The reported computational impact is substantial. On ITSDT-15k, pruning $123$8 reduces GFLOPs by $123$9 and improves $0.5\%$0 by $0.5\%$1; removing PAN further reduces parameters by $0.5\%$2M and GFLOPs by $0.5\%$3, for an overall compute reduction of about $0.5\%$4 and parameter reduction of about $0.5\%$5 relative to the pre-trimmed multi-head configuration, while maintaining or marginally improving accuracy (Atrash et al., 26 Sep 2025).

Training uses AdamW with learning rate $0.5\%$6 (Atrash et al., 26 Sep 2025). For multi-frame datasets such as ITSDT-15k and IRDST, inputs are resized to $0.5\%$7, YOLOv12n is initialized with COCO weights, training runs for $0.5\%$8 epochs, and batch size is $0.5\%$9. For single-frame datasets such as NUAA-SIRST and NUDT-SIRST, inputs are resized to $(<0.15)$0 and training runs for $(<0.15)$1 epochs (Atrash et al., 26 Sep 2025). The paper also reports a two-stage CA training procedure for the $(<0.15)$2 branch: the backbone and neck are frozen, the head is reinitialized with COCO weights, CA blocks are inserted only on the $(<0.15)$3 branch, and CA plus head are fine-tuned for $(<0.15)$4 epochs (Atrash et al., 26 Sep 2025).

The training hardware is a single NVIDIA A40 GPU with $(<0.15)$5 GB memory, while inference is reported on a single NVIDIA RTX 3080 Ti laptop GPU (Atrash et al., 26 Sep 2025). The fastest configuration is the ITSDT-15k P2-only model with PAN removed, which reaches up to about $(<0.15)$6 FPS; the NUAA-SIRST configuration with $(<0.15)$7 and partial PAN reaches about $(<0.15)$8 FPS (Atrash et al., 26 Sep 2025).

5. Empirical results and ablation structure

The reported benchmarks include ITSDT-15k and IRDST as multi-frame datasets, NUAA-SIRST and NUDT-SIRST as single-frame datasets, and IRDST-1k for cross-dataset validation (Atrash et al., 26 Sep 2025). The paper states that TY-RIST demonstrates state-of-the-art performance across four benchmarks and across $(<0.15)$9 different models, with cross-dataset validation on a fifth dataset (Atrash et al., 26 Sep 2025).

The most detailed ablations are on ITSDT-15k. The YOLOv12n baseline is reported at $2$0, Precision $2$1, Recall $2$2, $2$3, with $2$4M parameters and $2$5 GFLOPs (Atrash et al., 26 Sep 2025). Making the backbone stride-aware raises $2$6 by $2$7 to $2$8 and lifts $2$9 to $1$0 while increasing GFLOPs by $1$1 (Atrash et al., 26 Sep 2025). Replacing CIoU with NWD adds another $1$2 in $1$3 and raises $1$4 by $1$5 to $1$6, with best performance at $1$7 (Atrash et al., 26 Sep 2025). Adding the $1$8 head increases Precision by $1$9 and lifts $(P2)$0 to $(P2)$1, albeit with a substantial compute increase; the three CA blocks then add $(P2)$2 Recall for only $(P2)$3 GFLOPs (Atrash et al., 26 Sep 2025). After pruning and PAN removal, the model recovers efficiency while preserving or improving accuracy, yielding the headline overall gains of $(P2)$4 in $(P2)$5, $(P2)$6 in Precision, $(P2)$7 in Recall, and $(P2)$8 in $(P2)$9 relative to the YOLOv12n baseline (Atrash et al., 26 Sep 2025).

On ITSDT-15k, TY-RIST reports $+7.9\%$00 and $+7.9\%$01 (Atrash et al., 26 Sep 2025). In the comparison reported as SIRST, this is $+7.9\%$02 in $+7.9\%$03 over DNANet, although $+7.9\%$04 is $+7.9\%$05 lower; in the comparison reported as MIRST, it is $+7.9\%$06 in $+7.9\%$07 over STC, although $+7.9\%$08 is $+7.9\%$09 lower than Tridos (Atrash et al., 26 Sep 2025). The detector is also described as approximately $+7.9\%$10 lighter than DNANet and about $+7.9\%$11 lighter than Tridos in parameter count (Atrash et al., 26 Sep 2025).

On IRDST, TY-RIST is reported to achieve the best $+7.9\%$12 and $+7.9\%$13, surpassing the second-best by $+7.9\%$14 and $+7.9\%$15 among single-frame methods, and by $+7.9\%$16 and $+7.9\%$17 among multi-frame methods (Atrash et al., 26 Sep 2025). On NUAA-SIRST, with $+7.9\%$18 and partial PAN, it reports Precision $+7.9\%$19, Recall $+7.9\%$20, and $+7.9\%$21, corresponding to gains of $+7.9\%$22, $+7.9\%$23, and $+7.9\%$24 over prior bests (Atrash et al., 26 Sep 2025). On NUDT-SIRST, with only the $+7.9\%$25 head, it reports Precision $+7.9\%$26, Recall $+7.9\%$27, and $+7.9\%$28, exceeding prior bests by $+7.9\%$29, $+7.9\%$30, and $+7.9\%$31 (Atrash et al., 26 Sep 2025). Cross-dataset validation on IRDST-1k ranks second in Recall and $+7.9\%$32, and third in Precision, which the paper interprets as strong generalization (Atrash et al., 26 Sep 2025).

The paper also reports replication on YOLOv12s. The baseline YOLOv12s has $+7.9\%$33, $+7.9\%$34, $+7.9\%$35M parameters, and $+7.9\%$36 GFLOPs; after stride-aware changes, NWD $+7.9\%$37, the $+7.9\%$38 head, and CA$+7.9\%$39, the final pruned P2-only model with PAN removed reaches $+7.9\%$40 and $+7.9\%$41 (Atrash et al., 26 Sep 2025). This suggests that the design principles are not confined to a single backbone scale.

6. Terminological ambiguity and related uses of the name

The string “TY-RIST” is not a stable acronym across arXiv. In the infrared detection literature, it is explicitly the name of the YOLO-based IRSTD architecture described above (Atrash et al., 26 Sep 2025). In several other contexts, however, the same or similar string is either absent or interpreted differently.

In semi-supervised semantic segmentation, “The GIST and RIST of Iterative Self-Training for Semi-Supervised Segmentation” does not define or mention “TY-RIST”; the relevant term there is RIST, standing for Random Iterative Self-Training (Teh et al., 2021). In 3D semantic correspondence, RIST instead denotes a Rotation-Invariant 3D correspondence learner with Local Shape Transform (Park et al., 2024). In non-terrestrial cognitive NOMA networking, one paper states that “TY-RIST” corresponds to transmissive RIS $+7.9\%$42 in that work (Khan et al., 2024). A related cognitive RSMA paper does not define “TY-RIST” either, and uses TRIS for “transmissive reconfigurable intelligent surface transmitter” (Liu et al., 2023).

This suggests that the label is context-dependent rather than canonical. Within current arXiv usage, the most direct and explicit sense of TY-RIST is the IRSTD detector introduced in 2025 (Atrash et al., 26 Sep 2025); elsewhere, the term is either a naming variant, an interpretation of $+7.9\%$43-RIS/TRIS, or absent altogether (Khan et al., 2024).