Skip to main content
. Author manuscript; available in PMC: 2024 Feb 1.
Published in final edited form as: IEEE Trans Pattern Anal Mach Intell. 2023 Jan 6;45(2):2412–2429. doi: 10.1109/TPAMI.2022.3170559

Fig. 5.

Fig. 5.

Different ways of using encoders during inference. Top: encoders for multiple views are used and output representations are merged by combinations such as summation [10] or concatenation. Middle: only the main encoder [47] and the corresponding view are used during inference. Bottom: the given graph is directly input to the only encoder [48, 49] shared by all views to compute its representation.