Figure 2. Framework overview. Best viewed in color. We develop the high-level feature alignment into separate task spaces, by applying
the proposed Task-specific Inconsistency Alignment module to both the classification (green part) and localization (blue part) branches
of the baseline detector (gray part). In each branch, the behavioral inconsistency of multiple auxiliary predictors is optimized via the
corresponding inconsistency-aware loss, for essentially bridging the category-wise or boundary-wise margins between domains.