Asymmetric similarity improves traffic sign detection in YOLO benchmark tests

A YOLO-based traffic sign detector that treats feature matching as directional rather than symmetric improved detection accuracy on two public benchmarks while using slightly fewer parameters and fewer floating-point operations than its main control. The study developed Tversky-YOLO, a YOLO11m-scale system that introduces asymmetric similarity into both the residual backbone and the final classification head. Across five independently seeded runs, the model reached 96.8 ± 0.4% mean average precision at an intersection-over-union threshold of 0.5 and 82.1 ± 0.6% mAP averaged from IoU 0.5 to 0.95 on the German Traffic Sign Detection Benchmark. The controlled YOLO11m baseline achieved 95.1 ± 0.4% and 79.6 ± 0.5%, respectively, making the improvement across the stricter thresholds the primary reported result. The reported gains were statistically significant, with p = 0.003 and Cohen’s dz = 2.89 for GTSDB mAP@0.5:0.95.

The design starts from a mismatch between standard neural-network classifiers and the visual logic of traffic signs. Conventional one-by-one convolution classifiers calculate class scores with a dot product, which weights two kinds of mismatch equally: features activated in an image but absent from a class representation, and features expected by a class but missing from the image. For many traffic signs, however, the second mismatch can be more informative. A speed-limit sign and another circular sign may share a border and overall shape, while the presence or absence of a particular digit, arrow, colour, or prohibition symbol determines whether the detection is correct. The authors adapted the Tversky similarity index, originally developed to describe asymmetric judgments of similarity, so that learned parameters can penalize extraneous and missing activations separately. In the formulation used for classification, the alpha parameter weights extraneous activations in the input, while beta weights expected features missing from the input. Initializing beta above alpha gives the model a starting prior that missing diagnostic features are more consequential, although both parameters are subsequently learned.

Tversky-YOLO embeds this idea at two points in the YOLO11m architecture. The first is TverskyFusion, a lightweight gate applied to residual connections in the C2f backbone. For each block, non-negative versions of the input and transformed output are compared using an element-wise Tversky map. The resulting gate multiplies the residual contribution before it is added to the original input, suppressing transformed activations that conflict with the local input while preserving those that agree with it. The module uses shared learnable asymmetry parameters across backbone blocks and does not alter spatial resolution or channel counts. This mechanism differs from conventional channel or spatial attention because its gating signal is calculated directly from local asymmetric similarity rather than from a separately learned attention statistic. A control model used a learned one-by-one convolution followed by a sigmoid gate, giving the comparison a structurally similar alternative without the Tversky similarity map.

The second component, TverskyProjection, replaces the terminal one-by-one convolutional classifier at the P3, P4, and P5 detection scales. A learned feature bank projects each spatial feature vector into 256 non-negative activations, and every class has a non-negative prototype in that space. Tversky similarity compares the activation vector with each class prototype using separate alpha and beta penalties for features present in the activation but absent from the prototype and features expected by the prototype but absent from the activation. A log-odds transformation converts the resulting bounded similarity score into a logit that can be trained with the binary cross-entropy loss already used by YOLO. The final model learns alpha and beta separately for each class, adding only 86 scalar parameters for the 43 GTSDB categories. The bounding-box regression branch, objectness computation, and task-aligned assignment logic remain unchanged.

The proposed detector has essentially the same computational profile as its main baseline. YOLO11m has 231 layers, 20.11 million parameters, and 68.5 GFLOPs, compared with 20.07 million parameters and 67.9 GFLOPs for Tversky-YOLO. Both were trained with the same data splits, 640-by-640-pixel inputs, augmentation, optimizer, learning-rate schedule, batch size, and number of epochs. The similarity-based gate and prototype classifier therefore cannot be credited simply to a larger model. In measured inference on an NVIDIA A100, Tversky-YOLO recorded 41.65 ± 0.38 frames per second, compared with 38.10 ± 0.42 for the baseline, with p = 0.01. The researchers cautioned that these GPU measurements do not establish performance on representative embedded automotive hardware. The speed test used batch size one, averaged 1,000 forward passes after 200 warm-up passes, and therefore describes the reported laboratory configuration rather than a deployed vehicle system.

On the larger Tsinghua-Tencent 100K benchmark, Tversky-YOLO achieved 96.3 ± 0.5% mAP@0.5 and 74.4 ± 0.7% mAP@0.5:0.95. Its mean paired improvement over the matched YOLO11m baseline was 4.4 percentage points for the stricter metric, while mAP@0.5 rose by 3.5 points from a baseline value of 92.8%. The experiment used the standard split of 6,105 training and 3,071 test images and retained categories with at least 100 training instances. Per-class analysis showed higher average precision for 38 of the 45 retained categories, with a median improvement of 3.2 percentage points and an interquartile range of 1.8 to 5.1 points. No category accounted for more than 8.4% of the total gain, indicating that the aggregate result was not driven by a small number of frequent sign types.

Several of the largest TT100K gains occurred among categories that differ mainly in internal iconography. The “no overtaking” and “no overtaking for trucks” categories improved by 6.3 and 5.9 percentage points, respectively. Confusion analysis on GTSDB also linked the approach to its proposed mechanism. Relative to the baseline, the model reduced confusion between speed-30 and speed-50 signs by 61%, between no-passing and end-of-no-passing signs by 48%, and between give-way and stop signs by 44%. In an analysis of 50 cases in which the baseline misclassified a sign but Tversky-YOLO succeeded, EigenCAM maps from the modified model concentrated more sharply on diagnostic interiors such as numerals and prohibition symbols. Their mean overlap with manually annotated diagnostic elements was 0.71, compared with 0.48 for the baseline, a relative improvement of 48% in the reported localization measure. The study separately reported that learned asymmetry parameters formed patterns associated with different sign semantics, including heavier missing-feature penalties for several prohibitory signs and heavier extraneous-activation penalties for mandatory directional signs.

Controlled ablations show that the two components do not contribute equally across the datasets. On GTSDB, TverskyProjection with learned asymmetry reached 81.5% mAP@0.5:0.95, compared with 80.9% for a capacity-matched cosine-prototype classifier; the difference was reported as p = 0.04 with Cohen’s dz = 1.92. On TT100K, however, the asymmetric projection alone achieved 71.6%, essentially matching the symmetric control at 71.8%, a difference that was not statistically significant at p = 0.38. The larger TT100K advantage appeared when the projection was combined with TverskyFusion: the full model reached 74.4%, while the backbone module alone reached 72.8%, a difference reported as p = 0.009 with Cohen’s dz = 2.45. TverskyFusion also modestly outperformed a sigmoid-gated residual control on both datasets, 81.2% versus 80.8% on GTSDB and 72.8% versus 72.3% on TT100K. These results support a coupled interpretation in which similarity-aware feature refinement makes class-specific asymmetric matching more useful, rather than treating the two modules as independent additive improvements.

The study also tested whether the classification mechanism transfers beyond YOLO11m. Replacing the standard head of the smaller YOLOv8n model with either shared or per-class TverskyProjection improved performance over the conventional head by 1.2 and 1.9 percentage points, respectively, with both comparisons reported as significant at p < 0.01. The results nevertheless have clear boundaries. GTSDB contains only about 14 training images per class on average, potentially weakening prototype stability for the least frequent signs. Accuracy declines for signs smaller than 16 by 16 pixels in the resized inputs, and prototype comparison becomes unreliable when more than 60% of a sign is occluded. TT100K also has a long-tailed category distribution, making per-class metrics noisy when test sets are small. The experiments cover only traffic-sign benchmarks, and the authors state that the asymmetric weighting may need to be retuned for other fine-grained domains.

What makes the work distinctive is not a new detector family, but a change to the geometry of class comparison. The results suggest that a classifier can benefit from distinguishing a missing diagnostic feature from an unwanted extra feature, and that this distinction can also be used to regulate internal feature flow. Yet the evidence supports a carefully bounded conclusion: asymmetry produced a direct, measurable benefit for the GTSDB classifier, while on TT100K its clearest benefit emerged through interaction with similarity-aware backbone refinement. The approach therefore offers a promising design direction for fine-grained recognition, while future work must address extreme object sizes, occlusion robustness, embedded-hardware performance, and whether the learned directional pattern transfers beyond traffic signs. The study was conducted on public traffic-sign image benchmarks and does not establish road-safety benefits, autonomous-driving performance, or readiness for real-world deployment.

Subject of Research: Asymmetric Tversky similarity for fine-grained traffic sign detection with YOLO11m

Article Title: Asymmetric similarity aware object detection using Tversky YOLO for fine grained traffic sign recognition

Article References: Ben-Abbou, T., Omrani, H. E., Fazazy, K. E., Mahraz, M., Tairi, H., & Riffi, J. (2026). Asymmetric similarity aware object detection using Tversky YOLO for fine grained traffic sign recognition. Discover Artificial Intelligence, 6(1), Article 1081. https://doi.org/10.1007/s44163-026-02156-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02156-5

Keywords: traffic sign detection, Tversky similarity, YOLO11m, asymmetric classification, prototype learning, object detection, fine-grained recognition, GTSDB, TT100K, computer vision, Asymmetric, similarity