Table 3. Performance comparison of different feature fusion modules in the VisDrone2019 dataset.
Model | New detection head | BiFPN | CA | mAP/% | Parameter/million | t/s |
---|---|---|---|---|---|---|
YOLOv4 | 42.79 | 63.99 | 0.025 | |||
Version-1 | 43.73 | 52.54 | 0.023 | |||
Version-2 | √ | 52.13 | 44.62 | 0.024 | ||
Version-3 | √ | 44.04 | 52.61 | 0.023 | ||
Version-4 | √ | 43.93 | 52.58 | 0.026 | ||
Version-5 | √ | √ | 52.40 | 44.76 | 0.025 | |
Version-6 | √ | √ | 52.51 | 44.65 | 0.028 | |
Version-7 | √ | √ | 44.06 | 52.65 | 0.026 | |
YOLOv4-ours | √ | √ | √ | 52.76 | 44.79 | 0.028 |
Note:
The best results are in bold.