.png){kind=icon}
.png){kind=icon}
Return
Cross-domain object detection using minimized instance shift image-image translation
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-022-02643-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the rapid development of deep learning, the field of object detection has made a breakthrough. These techniques are data-driven, the performance of model largely depends on the dataset used for training. But the acquisition of a large-scale dataset is expensive and time-consuming. One alternative is to train the model on the expanded cross-domain dataset by image-image translation, while the existing image translation methods often fail to generalize mainly due to the instance shift which adverse to annotation reuse. In view of this problem, an image translation method based on improved CycleGAN is proposed in this paper. By the introducing of total variation loss and structural consistency loss to the full objective, the instance shifts between the source dataset and expanded cross-domain dataset is minimized, and the quality of the expanded cross-domain dataset has been greatly improved. The proposed approach is applied to the cross-domain detection in automatic driving scene, and the results demonstrate the robustness and superiority of the proposed method.
Keywords:
Cross-domain
Image-image translation
CycleGAN
Object detection
Journal
IF:
2.9
Papers:
4.6K
Citations:
6.5K
Organization
No organization information available