Fig. 1. Structure of DBL
Fig. 2. Structure of Res Unit
Fig. 3. Structure of Resn
Fig. 4. Structure of YOLOv3 DarkNet-53 baseline
Fig. 5. Contrast schematics between IOU and GIOU
Fig. 6. Schematic of Mixup interpolation
Fig. 7. Piecewise learning rate curve
Fig. 8. Cosine Decay learning rate curve
Fig. 9. Activation functions of Hard-Swish and Swish
Fig. 10. Effect picture of Cutout
Fig. 11. Effect pictures of ACDC. (a) VOC dataset; (b) KITTI 2D dataset
Fig. 12. Comparisonsof 20 classes of AP index on VOC dataset
Fig. 13. Comparisons of 7 classes of AP index on KITTI 2D dataset
Fig. 14. Comparisons of detection results of different models in traffic scenes. (a) Origin model; (b) +ACDC model; (c) +All model
YOLOv3 model | mAP /% | Δ /% | Time consuming /h |
---|
Origin | 76.55 | — | 22.72 | +Hard-Swish | 76.70 | +0.15 | 23.58 | +Swish | 76.66 | +0.11 | 25.07 |
|
Table 1. Comparisons of Hard-Swish and Swish activation function in mAP index
YOLOv3 model | mAP /% | Δ /% | FPS(416×416) /(frame/s) | Time consuming /h |
---|
Origin | 76.55 | — | | 22.72 | +Cosine Decay | 77.21 | +0.66 | | 22.73 | +ACDC | 78.19 | +1.64 | | 22.78 | +Mixup | 77.88 | +1.33 | 32.2(±0.5) | 22.77 | +Focal Loss | 81.02 | +4.47 | | 20.2 | +GIOU | 77.75 | +1.2 | | 22.75 | +Hard-Swish | 76.7 | +0.15 | | 23.58 | +All | 84.21 | +7.66 | | 22.48 |
|
Table 2. Comparisons of experiment indices on VOC dataset
YOLOv3 model | mAP /% | Δ /% | FPS(416×416) /(frame/s) | Time consuming /h |
---|
Origin | 81.92 | — | | 11.8 | +ACDC | 83.96 | +2.04 | 33.4(±0.5) | 11.82 | +All | 89.21 | +7.29 | | 11.66 |
|
Table 3. Comparisons of experiment indices on KITTI 2D dataset