DA³Net: Dual-Axis Aggregation Attention Network for Efficient Object Detection

This project is built based on Ultralytics.

Self-attention has demonstrated great potential for improving performance in object detection. However, the quadratic computational complexity with respect to the input size hinders its feasibility for high-resolution vision tasks in real-time scenarios. To effectively address this issue, we reinterpret self-attention as selective global context aggregation and propose a Dual-Axis Aggregation Attention (DA³) module. It achieves global perception with linear complexity by sequentially aggregating features along the distinct spatial axes, followed by depth-wise separable convolution to enhance feature selection. Based on this mechanism, we develop DA³Net, a lightweight backbone for efficient object detection. It also incorporates an enhanced local feature extraction module to improve the representation learning of fine-grained visual details guided by inductive bias compensation, which is an essential complement to long-range dependencies.

Extensive experiments validate the effectiveness of the proposed approach. On MS COCO, DA³Net-N achieves 41.7 AP, exceeding YOLO11-N by 2.2 AP. Ablation studies on the PASCAL VOC further demonstrate the effectiveness and efficiency of each design component, with DA³Net-N attaining 63.17 mAP, which is 2.4 mAP higher than YOLO11-N. On a RTX A5000 GPU, it runs at 420.63 FPS in TensorRT format—over 12x faster than standard self-attention—demonstrating strong real-time capability. Moreover, generalization experiments on underwater object detection further validate the robustness and adaptability of the proposed approach.

Results

Comparison with popular state-of-the-art real-time object detectors on the MS COCO 2017 dataset.

| Method | Param.(M) | FLOPs(G) | AP_50:95 | AP₅₀ | AP₇₅ | |--------------------------|---------------|--------------|------------------------|---------------------|---------------------| | YOLOv6-3.0-N | 4.7 | 11.4 | 37.0 | 52.7 | - | | Gold-YOLO-N | 5.6 | 12.1 | 39.9 | 56.0 | 43.3 | | YOLOv8-N | 3.2 | 8.7 | 37.7 | 52.5 | 40.6 | | YOLOv10-N | 2.3 | 6.7 | 38.5 | 63.7 | 61.6 | | YOLO11 | 2.6 | 6.5 | 39.5 | 55.1 | 42.8 | | DA³Net-N | 2.6 | 6.9 | 41.7 | 58.2 | 44.9 |

| Method | Param.(M) | FLOPs(G) | AP_50:95 | AP₅₀ | AP₇₅ | |--------------------------|---------------|--------------|------------------------|---------------------|---------------------| | YOLOv6-3.0-S | 18.5 | 45.3 | 44.3 | 61.2 | - | | Gold-YOLO-S | 21.5 | 46.0 | 45.4 | 63.5 | 50.4 | | YOLO-MS-XS | 4.5 | 17.4 | 43.4 | 60.4 | 47.6 | | YOLO-MS-S | 8.1 | 31.2 | 46.2 | 63.7 | 50.5 | | YOLOv8-S | 11.2 | 28.6 | 45.0 | 61.8 | 48.7 | | YOLOv9-S | 7.1 | 26.4 | 46.7 | 63.4 | 50.7 | | RT-DETR-R18 | 20.0 | 60.0 | 46.5 | 63.8 | - | | YOLOv10-S | 7.2 | 21.6 | 46.3 | 62.9 | 50.2 | | YOLO11-S | 9.5 | 21.7 | 46.9 | 63.8 | 50.7 | | DA³Net-S | 9.5 | 22.8 | 48.5 | 65.6 | 52.7 |

| Method | Param.(M) | FLOPs(G) | AP_50:95 | AP₅₀ | AP₇₅ | |--------------------------|---------------|--------------|------------------------|---------------------|---------------------| | YOLOv6-3.0-M | 34.9 | 85.8 | 49.1 | 66.1 | - | | Gold-YOLO-M | 51.3 | 87.5 | 49.8 | 67.0 | - | | YOLOv8-M | 25.9 | 78.9 | 50.3 | 67.2 | 54.7 | | RT-DETR-R34 | 36.0 | 100.0 | 48.9 | 66.8 | - | | RT-DETRv2-R34 | 36.0 | 100.0 | 49.9 | 67.5 | - | | YOLOv9-M | 20.0 | 76.3 | 51.4 | 68.1 | 56.1 | | YOLOv10-M | 15.4 | 59.1 | 51.1 | 68.1 | 55.8 | | YOLO11-M | 20.1 | 68.0 | 51.5 | 68.5 | 55.7 | | DA³Net-M | 18.3 | 68.8 | 52.3 | 69.7 | 57.0 |

| Method | Param.(M) | FLOPs(G) | AP_50:95 | AP₅₀ | AP₇₅ | |--------------------------|---------------|--------------|------------------------|---------------------|---------------------| | YOLOv6-3.0-L | 59.6 | 150.7 | 51.8 | 69.2 | - | | Gold-YOLO-L | 75.1 | 151.7 | 51.8 | 68.9 | - | | YOLOv8-L | 43.7 | 165.2 | 53.0 | 69.8 | 57.7 | | RT-DETR-R50 | 42.0 | 136.0 | 53.1 | 71.3 | - | | RT-DETRv2-R50 | 42.0 | 136.0 | 53.4 | 71.6 | - | | YOLOv9-C | 25.3 | 102.1 | 53.0 | 70.2 | 57.8 | | YOLOv10-B | 19.1 | 92.0 | 52.5 | 69.6 | 57.2 | | YOLOv10-L | 24.4 | 120.3 | 53.2 | 70.1 | 58.1 | | YOLO11-L | 25.3 | 86.9 | 53.3 | 70.1 | 58.2 | | DA³Net-L | 27.7 | 106.9 | 53.9 | 70.6 | 58.6 |

| Method | Param.(M) | FLOPs(G) | AP_50:95 | AP₅₀ | AP₇₅ | |--------------------------|---------------|--------------|------------------------|---------------------|---------------------| | YOLOv8-X | 68.2 | 257.8 | 54 | 71 | 58.8 | | RT-DETR-R101 | 76 | 259 | 54.3 | 72.7 | - | | RT-DETRv2-R101 | 76 | 259 | 54.3 | 72.8 | - | | YOLOv10-X | 29.5 | 160.4 | 54.4 | 71.3 | 59.3 | | YOLO11-X | 56.9 | 194.9 | 54.6 | 71.6 | 59.5 | | DA³Net-X | 61.7 | 234.8 | 55.1 | 71.8 | 60.0 |

Training

```python

from ultralytics import YOLO

Load a model

model = YOLO("yolo113n.yaml") # yolo113 is the alias of DA3Net

Train the model

trainresults = model.train( data="coco.yaml", epochs=500, # number of training epochs batch=128, imgsz=640, # training image size scale=0.5, # N:0.5, S:0.9; M:0.9; L:0.9; X:0.9 mosaic=1.0, mixup=0.0, # N:0.0, S:0.05; M:0.15; L:0.15; X:0.2 copypaste=0.1, # N:0.1, S:0.15; M:0.4; L:0.5; X:0.6 device=[0, 1], # device to run on, i.e. device=0 or device=0,1,2,3 or device=cpu project="runs/yolo113_coco", name="113n" )

```

Validation

```python from ultralytics import YOLO

model = YOLO('runs/yolo113_coco/113n/weights/best.pt') print(model) model.info()

metrics = model.val(data='coco.yaml', batch=1)

print("mAP50-95:", metrics.box.map) # mAP50-95 print("mAP50:", metrics.box.map50) # mAP50 print("mAP75:",metrics.box.map75) # mAP75 print("list of mAP50-95 for each category:", metrics.box.maps) # list of mAP50-95 for each category ```

Citation

If you find this project useful in your research, please consider cite:

Coming soon