Table 6 The influence of using different attention at different stages of the backbone on the model performance evaluation indicators.
From: Object detection model design for tiny road surface damage
Feature Maps | Attention | mAP50(%) | mAP50:95(%) | GFLOPs |
|---|---|---|---|---|
C2/C3/C4 | – | 66.9 | 35.5 | 24.4 |
C4 | ECA | 67.0(+ 0.1) | 35.8(+ 0.3) | 24.4(+ 0.0) |
GAM37 | 67.2(+ 0.3) | 35.9(+ 0.4) | 29.7(+ 5.3) | |
CAM37 | 67.1(+ 0.2) | 35.6(+ 0.1) | 24.7(+ 0.3) | |
PAM37 | 66.1(− 0.8) | 35.4(− 0.1) | 24.7(+ 0.3) | |
TripleAttention | 66.0(− 0.9) | 34.9(− 0.6) | 24.5(+ 0.1) | |
C3 | ECA | 66.6(− 0.3) | 35.0(− 0.5) | 24.4 |
GAM | 66.0(− 0.9) | 35.2(− 0.3) | 29.7(+ 5.3) | |
CAM | 66.5(− 0.4) | 35.4(− 0.1) | 24.6(+ 0.2) | |
PAM | 65.5(− 1.4) | 34.9(− 0.6) | 24.7(+ 0.3) | |
TripleAttention | 67.2(+ 0.3) | 35.8(+ 0.3) | 24.5(+ 0.1) | |
CBAM38 | 65.2(− 1.7) | 34.6(− 0.9) | 24.5(+ 0.1) | |
CoTAttention39 | 67.0(+ 0.1) | 35.6(+ 0.1) | 26.3(+ 1.9) | |
Nonlocal40 | 66.6(− 0.3) | 35.6(+ 0.1) | 25.3(+ 0.9) | |
MSCA41 | 66.0(− 0.9) | 35.1(− 0.4) | 24.7(+ 0.3) | |
SeaAttention42 | 67.0(+ 0.1) | 35.5(–) | 47.4(+ 23.0) | |
LSKblock | 67.1(+ 0.2) | 35.8(+ 0.3) | 25.2(+ 0.8) | |
C2 | GAM | 66.4(− 0.5) | 35.5(–) | 29.7(+ 5.3) |
PAM | 66.4(− 0.5) | 35.4(− 0.1) | 24.7(+ 0.3) | |
TripleAttention | 66.9(–) | 35.6(+ 0.1) | 24.5(+ 0.1) | |
CBAM | 66.3(− 0.6) | 35.3(− 0.2) | 24.5(+ 0.1) | |
CoTAttention | 66.6(− 0.3) | 35.9(+ 0.4) | 26.3(+ 1.9) | |
Nonlocal | 65.9(− 1.0) | 35.1(− 0.4) | 25.3(+ 0.9) | |
MSCA | 67.2(+ 0.3) | 35.5(–) | 24.8(+ 0.4) | |
SeaAttention | 65.5(− 1.4) | 35.1(− 0.4) | 40.9(+ 16.5) | |
LSKblock | 67.2(+ 0.3) | 35.9(+ 0.4) | 25.3(+ 0.9) |
