Table 2.
Performance comparison of competing methods using different supervision on the 2020 version of the GWHD dataset and the latest 2021 version
| Method (years) | Supervision | GWHD_2020 | GWHD_2021 | ||
|---|---|---|---|---|---|
| MAE | RMSE | MAE | RMSE | ||
| Faster R-CNN (2015) [35] | Box | 3.27 | 4.49 | 5.46 | 9.76 |
| SSD (2016) [33] | Box | 3.97 | 5.44 | 7.19 | 12.08 |
| DETR (2022) [36] | Box | 3.76 | 4.91 | 8.41 | 16.20 |
| YOLOv8 (2023) [34] | Box | 4.39 | 6.17 | 8.96 | 13.43 |
| MCNN (2016) [37] | Point | 5.05 | 6.47 | 7.59 | 10.17 |
| CSRNet (2018) [31] | Point | 3.87 | 5.01 | 5.88 | 8.13 |
| ASD (2019) [38] | Point | 5.86 | 7.55 | / | / |
| SPN (2019) [39] | Point | 5.91 | 7.73 | / | / |
| WheatNet (2022) [14] | Point | 3.85 | 5.19 | / | / |
| TransCrowd-Token (2022) [21] | Count | 15.87 | 19.02 | 11.29 | 14.39 |
| TransCrowd-GAP (2022) [21] | Count | 12.23 | 15.02 | 10.09 | 12.85 |
| CSNet (ours) | Count | 2.94 | 3.88 | 3.87 | 5.60 |