Revisiting the Receptive Field of Conv-GRU in DROID-SLAM

This work focuses on improving the Conv-GRU-based optical flow update within a DROID-SLAM framework. Prior optical flow models typically follow a UNet or coarse-to-fine architecture in order to extract long-range cross-correlation and context cues. This helps flow estimation in the presence of large motion and challenging image regions, e.g., textureless regions. We propose modifications to the Conv-GRU module which follows the rationale of these prior models by integrating (Atrous) Spatial Pyramid Pooling and global self-attention into the Conv-GRU block. By enlarging the receptive field through the aforementioned modifications, the model is able to integrate information from a larger context window, thus improving the robustness even when given inputs that comprise challenging image regions. We show empirically through extensive experiments the gain in accuracy through these modifications.