Author Affiliations
1College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410000, Hunan , China2Hunan Key Laboratory of Image Measurement and Vision Navigation, Changsha 410000, Hunan , Chinashow less
Fig. 1. Overall architecture of AO2DINO
Fig. 2. Multi-scale rotated deformable attention module
Fig. 3. Comparison of attention heatmaps between (left) and ReDet (right)
Fig. 4. Self-adaption assigner. (a) Positive and negative assigner of (left) and DINO (right); (b) overlapping of rotated boxes
Fig. 5. Calculation of Rotated IoU
Fig. 6. Periodicity of angle. (a) Ideal representation of bounding boxes; (b) the predicted angle differs from the ideal angle by 90; (c) the predicted angle differs from the ideal angle by 180
Fig. 7. Edge exchangeability
Fig. 8. Principle of KFIoU
Fig. 9. Comparison of test results of different methods on DOTAv1.0 dataset
Fig. 10. Adaptability of AO2DINO on DIOR-R dataset
Fig. 11. Comparison of dense small object detection performance on DOTAv1.0 dataset
| Configuration | Model |
|---|
| Operating system | Ununtu 20.0.4 | | GPU | NVIDIA GeForce RTX-4080Ti GPU | | Hardware configuration | i9-10920X | | Environment | Python 3.8，PyTorch1.7.1，CUDA11.2 |
|
Table 1. Experimental software and hardware configuration
| Category | One-stage | Two-stage | DETR-like |
|---|
| Rotated RetinaNet（3×） | R3Det （3×） | Rotated FCOS（3×） | Rotated Faster R-CNN（3×） | ReDet （3×） | Rotated D-DETR（3×） | AO2DETR （3×） | ARS-DETR （3×） | AO2DINO （1×） | AO2DINO （3×） |
|---|
| mAP | 69.23 | 73.40 | 72.45 | 73.96 | 74.03 | 63.42 | 70.91 | 73.79 | 72.16 | 74.07 | | PL | 87.33 | 89.24 | 88.52 | 89.09 | 88.94 | 78.95 | 87.99 | 86.61 | 86.33 | 86.87 | | BD | 78.91 | 83.32 | 77.54 | 78.28 | 78.07 | 68.64 | 79.46 | 77.26 | 76.79 | 81.91 | | BR | 46.45 | 48.03 | 47.06 | 48.93 | 51.19 | 32.57 | 45.74 | 48.84 | 49.52 | 48.25 | | GTF | 69.81 | 72.52 | 63.78 | 71.54 | 72.76 | 55.17 | 66.64 | 66.76 | 63.43 | 72.90 | | SV | 67.72 | 77.52 | 80.42 | 74.01 | 74.26 | 72.53 | 78.90 | 78.38 | 77.43 | 79.92 | | LV | 62.34 | 76.72 | 80.50 | 74.99 | 78.08 | 57.77 | 73.90 | 78.96 | 62.83 | 63.24 | | SH | 73.59 | 86.48 | 87.34 | 85.90 | 87.44 | 73.71 | 73.30 | 87.40 | 84.54 | 85.87 | | TC | 90.85 | 90.89 | 90.39 | 90.84 | 90.84 | 88.36 | 90.40 | 90.61 | 90.12 | 88.23 | | BC | 82.79 | 82.33 | 77.83 | 86.87 | 80.79 | 75.46 | 80.55 | 82.76 | 83.92 | 82.89 | | ST | 79.37 | 83.51 | 84.13 | 85.03 | 78.59 | 79.34 | 85.89 | 82.19 | 84.82 | 86.87 | | SBF | 59.62 | 60.96 | 55.45 | 57.97 | 60.85 | 45.36 | 55.19 | 54.02 | 55.94 | 61.17 | | RA | 61.89 | 63.09 | 65.84 | 69.74 | 64.22 | 53.78 | 63.62 | 62.61 | 67.22 | 65.97 | | HA | 65.01 | 67.58 | 65.02 | 68.10 | 76.84 | 52.94 | 51.83 | 72.64 | 68.11 | 65.60 | | SP | 67.76 | 69.27 | 72.77 | 71.28 | 72.79 | 66.35 | 70.15 | 72.80 | 71.89 | 77.39 | | HC | 44.95 | 49.50 | 49.17 | 56.88 | 54.85 | 50.38 | 60.04 | 64.96 | 75.40 | 77.13 |
|
Table 2. Comparison of different models on the DOTAv1.0 dataset
| Category | One-stage | Two-stage | DETR-like |
|---|
| Rotated RetinaNet（3×） | R3Det （3×） | Rotated FCOS（3×） | GWD （3×） | KLD （3×） | Rotated Faster R-CNN（3×） | ReDet （3×） | ARS-DETR （3×） | AO2DINO （1×） | AO2DINO （3×） |
|---|
| mAP | 54.83 | 61.91 | 63.21 | 60.31 | 64.63 | 63.41 | 63.81 | 65.90 | 60.54 | 65.94 | | APL | 59.54 | 62.55 | 62.31 | 69.68 | 66.52 | 63.07 | 63.22 | 65.82 | 63.93 | 68.78 | | APO | 25.03 | 43.44 | 42.18 | 28.83 | 46.80 | 40.22 | 44.18 | 53.40 | 42.21 | 48.83 | | BF | 70.08 | 71.72 | 75.34 | 74.32 | 71.76 | 71.89 | 72.11 | 74.22 | 73.24 | 74.32 | | BC | 81.01 | 81.84 | 81.32 | 81.49 | 81.43 | 81.36 | 81.26 | 81.11 | 83.57 | 84.49 | | BR | 28.26 | 36.49 | 39.26 | 29.62 | 40.81 | 39.67 | 43.83 | 42.13 | 40.39 | 41.62 | | CH | 72.02 | 72.63 | 74.89 | 72.67 | 78.25 | 72.51 | 72.72 | 76.23 | 63.65 | 72.67 | | ESA | 55.35 | 79.50 | 77.42 | 76.45 | 79.23 | 79.19 | 79.10 | 82.24 | 64.91 | 76.45 | | ETS | 56.77 | 64.41 | 68.67 | 63.14 | 66.63 | 69.45 | 69.78 | 71.52 | 68.98 | 69.14 | | DAM | 21.26 | 27.02 | 26.00 | 27.13 | 29.01 | 26.00 | 28.45 | 38.90 | 33.45 | 34.13 | | GF | 65.70 | 77.36 | 73.94 | 77.19 | 78.68 | 77.93 | 78.69 | 75.91 | 71.24 | 71.19 | | GTF | 70.28 | 77.17 | 78.73 | 78.94 | 80.19 | 82.28 | 77.18 | 77.91 | 77.03 | 78.94 | | HA | 30.52 | 40.53 | 41.28 | 39.11 | 44.88 | 46.91 | 48.24 | 33.03 | 42.67 | 43.11 | | OP | 44.37 | 53.33 | 54.19 | 42.18 | 57.23 | 53.90 | 56.81 | 57.02 | 66.65 | 66.18 | | SH | 77.02 | 79.66 | 80.61 | 79.10 | 80.91 | 81.03 | 81.17 | 84.82 | 85.43 | 86.10 | | STA | 59.01 | 69.22 | 66.92 | 70.41 | 74.17 | 75.77 | 69.17 | 69.71 | 69.80 | 70.41 | | STO | 59.39 | 61.10 | 69.17 | 58.69 | 68.02 | 62.54 | 62.73 | 72.20 | 62.34 | 62.69 | | TC | 81.18 | 81.54 | 87.20 | 81.52 | 81.48 | 81.42 | 81.42 | 80.33 | 72.98 | 81.66 | | TS | 38.43 | 52.18 | 52.31 | 47.78 | 54.63 | 54.50 | 54.90 | 58.91 | 54.55 | 55.78 | | VE | 39.10 | 43.57 | 47.08 | 44.47 | 47.80 | 43.17 | 44.04 | 51.52 | 49.80 | 50.47 | | WM | 61.58 | 64.13 | 65.21 | 62.36 | 64.41 | 65.73 | 66.37 | 70.73 | 68.21 | 69.36 |
|
Table 3. Comparison of different models on the DIOR-R dataset
| Method | Epoch | DOTAv1.0 | DIOR-R |
|---|
| SH | SV | SH | VE | WM |
|---|
| R3Det | 3× | 86.48 | 77.52 | 79.66 | 43.57 | 64.13 | | ReDet | 3× | 87.44 | 74.26 | 81.17 | 44.04 | 66.37 | | ARS-DETR | 3× | 87.40 | 78.38 | 84.82 | 51.52 | 70.73 | | AO2DINO | 3× | 85.87 | 79.92 | 86.10 | 50.47 | 69.36 | | AO2DINO | 1× | 84.54 | 77.43 | 85.43 | 49.80 | 68.21 | | AO2DINO-ms | 1× | 88.90 | 79.98 | 87.57 | 50.66 | 70.68 |
|
Table 4. mAP of dense small target detection on DOTAv1.0 and DIOR-R datasets
| CDN | MS-RDA | SAA | KFIoU | AP50 /% | AP75 /% |
|---|
| √ | | | | 67.12 | 33.35 | | √ | √ | | | 68.90（+1.78） | 38.70（+5.35） | | √ | | √ | | 71.06（+3.94） | 40.15（+6.80） | | √ | | | √ | 70.29（+3.17） | 36.65（+3.30） | | √ | √ | √ | √ | 72.16（+5.04） | 41.80（+8.45） |
|
Table 5. Ablation experiments of AO2DINO's component on DOTAv1.0 dataset
| Baseline | Scale | ResNet50 | Swin-T | AP50 /% | AP75 /% |
|---|
| AO2DINO | 4 scale | √ | | 72.16 | 41.80 | | √ | 72.50 | 42.10 | | 5 scale | √ | | 72.54 | 41.73 | | √ | 72.68 | 42.21 | | multi-scale | √ | | 75.77 | 44.29 |
|
Table 6. Comparative experiment of AO2DINO with different scales on DOTAv1.0 dataset
| Loss function | DOTAv1.0 | DIOR-R |
|---|
| L1 loss | 67.12 | 53.50 | | GWD | 70.01 | 55.56 | | KLD | 69.82 | 55.91 | | KFIoU | 70.29 | 56.02 |
|
Table 7. AP50 of different loss functions on DOTAv1.0 dataset and DIOR-R dataset
