Figure 4. Overview of our method. Motivated by the observation that certain feature channels learned from the source domain negatively
impact cross-domain performance, we propose suppressing and pruning sensitive channels while adapting invariant ones. We introduce
sensitivity-guided channel pruning, which quantifies the importance of each feature channel based on its image-level and instance-level
sensitivity to domain discrepancies. We then apply a weighted sparsity loss to the learnable parameters, promoting the pruning of channels
with higher sensitivity based on the predefined threshold and ratio. To prevent the loss of potentially useful channels from early pruning,
we introduce a stochastic channel reactivation mechanism.