Fig. 1. Framework of M2Det
Fig. 2. Structure of different modules in FFM. (a)FFMv1; (b)FFMv2
Fig. 3. Structure of different models. (a) FPN; (b) RFP
Fig. 4. Structure of different modules. (a) CAM; (b) SAM
Fig. 5. EM2Det model
Fig. 6. Structure of feature fusion enhancement module
Fig. 7. Detail of feature fusion enhancement module
Fig. 8. Details of recursive modules for U-shaped network
Fig. 9. Residual side attention module
Fig. 10. Some images in SIXray OD dataset. (a) Gun; (b) wrench; (c) knife; (d) pliers; (e) scissors
Fig. 11. Visualizations in different situations. (a) Scissors; (b) pliers; (c) gun; (d) wrench; (e) knife
Number of TUMs | Parameters /Mbit | mAP /% |
---|
2 | 182.05 | 73.34 | 4 | 227.88 | 73.99 | 6 | 293.61 | 76.89 | 8 | 331.62 | 79.03 | 10 | 383.78 | 79.62 |
|
Table 1. Index under different number of TUMs
Structure | Integration of pool 5 | mAP /% |
---|
Baseline(conv 4_3+pool 5) | × | 79.03(0) | Three layers (pool 3+conv 4_3+pool 5) | × | 81.94(2.91↑) | Three layers (pool 3+conv 4_3+pool 5) | √ | 82.82(3.79↑) | Four layers (pool 2+pool 3+conv 4_3+pool 5) | × | 81.13(2.10↑) | Four layers (pool 2+pool 3+conv 4_3+pool 5) | √ | 82.05(3.02↑) |
|
Table 2. Experiment effect of feature fusion enhancement module
Structure | Parameters /Mbit | Residual edge attention module | mAP /% |
---|
8 TUM | 334.50 | × | 82.82 | 1 RTUM +7 TUM | 356.67 | × | 83.61(0.79↑) | 8 RTUM | 514.89 | × | 84.09(1.27↑) | 8 RTUM+CBAM | 517.89(3.00↑) | × | 85.12(2.30↑) | √ | 85.40(2.58↑) | 8 RTUM+ECA | 511.86(3.00↓) | × | 84.86(2.04↑) | √ | 85.02(2.20↑) |
|
Table 3. Experimental effects of U-shaped network recursive module and residual edge attention module
Framework | Accuracy /% | mAP /% |
---|
Gun | Pilers | Scissors | Knife | Wrench |
---|
SSD | 92 | 72 | 66 | 55 | 52 | 70.9 | Faster RCNN | 93 | 75 | 71 | 65 | 67 | 74.2 | YOLOv3 | 92 | 71 | 64 | 53 | 51 | 66.2 | M2Det | 94 | 83 | 76 | 74 | 68 | 79.0 | M2Det+EFFM | 96 | 86 | 82 | 77 | 73 | 82.8 | M2Det+EFFM+RTUM | 96 | 88 | 82 | 78 | 76 | 84.0 | EM2Det | 98 | 89 | 84 | 79 | 77 | 85.4 |
|
Table 4. Accuracy of different detection models