Viewpoint Integration and Registration with Vision Language Foundation Model for Image Change Understanding
只有fused adapter image encoder, viewpoint registration flow, semantic emphasizing module, 和 fully connected layer 训练,其他参数冻结。
Fused Adapter Image Encoder
adapter:
fused adapter:
Viewpoint Registration Flow and Semantic Emphasizing
Viewpoint Registration Flow:
conv1是1x1 ; conv是3x3
,双线性插值
Semantic Emphasizing:
结果展示: