Imbalance dataset is the major problem we all will face in the process of developing deep learning model. There were many approaches to solve this very problem such as heuristic data sampling and modifying loss function for model training. In order to find the solution, we chose Foggy Cityscapes dataset for the experiment since this dataset has imbalance object class distribution. We proposed a method to solve imbalance dataset namely instance level downsampling as an extension of traditional downsampling method. The algorithm of this method will selectively keep and drop certain image in the dataset by evaluating the majority and minority object class proportion inside a single image. After comparing the model evaluation using Mean Average Precision (mAP) metric, the model which was trained with balanced dataset has more balanced knowledge or less biased across the object classes of interest.