I tried to replace the Segment-Anything (SAM) encoder with the DINOv2 encoder. I first use a 1*1 Conv to reduce the dimension of DINOv2 embedding (1024 dimension) to the same as SAM embedding (256 dimension), then the embedding is sent to SAM decoder for output. However, the segmentation result is noisy (containing some regular noise which is weird?). Could anyone help me out? Thanks! please see the segmentation mask here
I first use a 1*1 Conv to reduce the dimension of DINOv2 embedding (1024 dimension) to the same as SAM embedding (256 dimension), then the embedding is sent to SAM decoder for output. However, the segmentation result is noisy (containing some regular noise which is weird?). Could anyone help me out? Thanks! please see the segmentation mask here