Efficient Mirror Detection via Multi-level Heterogeneous Learning

We present HetNet (Multi-level \textbf{Het}erogeneous \textbf{Net}work), ahighly efficient mirror detection network. Current mirror detection methodsfocus more on performance than efficiency, limiting the real-time applications(such as drones). Their lack of efficiency is aroused by the common design ofadopting homogeneous modules at different levels, which ignores the differencebetween different levels of features. In contrast, HetNet detects potentialmirror regions initially through low-level understandings (\textit{e.g.},intensity contrasts) and then combines with high-level understandings(contextual discontinuity for instance) to finalize the predictions. To performaccurate yet efficient mirror detection, HetNet follows an effectivearchitecture that obtains specific information at different stages to detectmirrors. We further propose a multi-orientation intensity-based contrastedmodule (MIC) and a reflection semantic logical module (RSL), equipped onHetNet, to predict potential mirror regions by low-level understandings andanalyze semantic logic in scenarios by high-level understandings, respectively.Compared to the state-of-the-art method, HetNet runs 664$\%$ faster and drawsan average performance gain of 8.9$\%$ on MAE, 3.1$\%$ on IoU, and 2.0$\%$ onF-measure on two mirror detection benchmarks.