Research Article
Unsupervised Image-Generation Enhanced Adaptation for Object Detection in Thermal Images
Table 2
Ablation study for each proposed component.
| Image trans | Int-Inv1 | R-A2 | Bike | Car | Car_stop | Color_cone | Person | mAP |
| No | — | | 1.5 | 0.7 | 0.5 | 0 | 4.1 | 1.4 | No | — | √ | 1.04 | 3.1 | 0.71 | 0.03 | 39.0 | 8.8 | Gray | | | 0.94 | 4.1 | 1.2 | 2.03 | 18.3 | 5.3 | Gray | √ | | 3.9 | 12.2 | 5.0 | 7.5 | 52.5 | 16.2 | Gray | | √ | 1.5 | 2.1 | 2.5 | 7.2 | 50.7 | 12.8 | Gray | √ | √ | 3.0 | 7.3 | 4.7 | 17.0 | 61.2 | 18.6 | CycleGAN | | | 12.8 | 16.4 | 3.5 | 4.0 | 26.1 | 12.6 | CycleGAN | √ | | 18.7 | 31.6 | 6.0 | 6.6 | 49.0 | 22.4 | CycleGAN | | √ | 13.3 | 14.4 | 6.3 | 7.3 | 41.5 | 16.6 | CycleGAN | √ | √ | 20.3 | 30.8 | 11.7 | 12.9 | 56.9 | 26.5 |
|
|
Note. 1Int-Inv indicates intensity inversion; 2R-A means readaptation.
|