Fig. 1.

Illustration of proposed NIA-Network architecture. For easier understanding, we utilize two functions ($f_{\mathit{tp}}(·)$ and $f_{\mathit{dc}}(·)$) to illustrate our NIA-Network. The function $f_{\mathit{tp}}(·))$ contains an extractor (a vanilla ResNet50), NI-Module and a vanilla Faster R-CNN, and it is a feed-forward Conv-based architecture. The function $f_{\mathit{dc}}(·))$ indicates the domain adaptation architecture, and it contains a domain classifier and one gradient reversal layer (GRL). The function $f_{\mathit{dc}}(·))$ ensures that representations over the two domains are similar which enables to transfer labels from source domain to target domain. Our NI-Module is shown in Section 4.1 and domain adaptation architecture is introduced in Section 4.2.