Closed XIE-ZJU closed 8 months ago
Hi, (1) The shape of "Dense Language Features" is [512, H, W], achieved by summing the [scale, D, H, W] tensor along the dimension 0 and then normalizing it. (2) In order to ensure consistency in the magnitude of high-dimensional features, we normalize their length to 1.
OK, thanks for your quick reply! I get it!
Hi, thanks for your great job of this excellent work. Recently, I am working on the similar region of you. In the paper, you said "We aggregate features from all layers and normalize them to produce the final language embeddings." I have few questions: (1) I extracted the unnormalized embeddings following 3DOVS and get a [scale, D, H, W] tensor ([3, 512, H, W] in practice), and is this the "Dense Language Features" you said in the paper? (2) Whether the type of the normalization matters? Is it min-max, z-score or others? I'm sincerely looking for your reply. Thanks again!