Unsupervised Domain Adaptive Object Detection Using Forward-Backward Cyclic Adaptation

We present a novel approach to perform the unsupervised domain adaptation for object detection through forward-backward cyclic (FBC) training. Recent adversarial training based methods have shown their effectiveness on minimizing discrepancy via marginal feature distributions alignment. However, aligning does not guarantee alignment of class conditional distributions. This limitation is more evident when adapting detectors as larger compared image classification task, e.g. various number objects exist in one and majority content an background. motivates us learn invariance category level semantics gradient Intuitively, if gradients two domains point similar directions, then learning can improve that another domain. To achieve alignment, we propose Forward-Backward Cyclic Adaptation, which iteratively computes from source target backward hopping forward passing. In addition, align low-level features holistic color/texture detector performs well both ideal As such, each cycle, diversity enforced by maximum entropy regularization penalize confident source-specific minimum intrigue target-specific learning. Theoretical analysis process provided, extensive experiments challenging cross-domain datasets superiority our over state-of-the-art.