Federated Online Adaptation for Deep Stereo

This document supplements the CVPR 2024 paper "Federated Online Adaptation for Deep Stereo". It provides additional implementation details and deeper insights into the results reported in the main paper. Implementation Details MADNet 2 Architecture MADNet 2 is implemented on top of MADNet [55]. Specifically, it is made of two, shared feature extractors and a set of five shallow disparity decoders. These modules are assembled to implement coarse-to-fine processing, as shown in Fig. 5 . Figure 5. MADNet 2 architecture. Given a stereo pair, a set of multi-scale features is extracted by means of two feature extractors with shared weights. Starting from the lowest resolutioni.e., 1 64 -correlation scores are computed and sampled by means of the all-pair correlation module and lookup operator from [30]. Sampled scores and image features are processed by a disparity decoder, which predicts an initial disparity map at 1 64 resolution. This latter is upsampled and used by the look operator working on the correlation volume at 1 32 resolution, then a second decoder predicts a refined disparity map at 1 32 resolution. This process is repeated up to 1 4 resolution. There, the final prediction is bilinearly upsampled to the original resolution.