Data augmentation is being widely used to enrich the datasets and enhance the performance of neural network for classification and detection. However most of the recent works focus only on the augmentation for classification. A technique for detection augmentation by template blending has been introduced in the literature. The limitation of blending technique is an extra polygon shape of each object needed to blend with the scene. In this paper we investigate the effect of the geometric transformations for detection augmentation on the Malaysian Traffic Sign Detection (MTSD) datasets. We propose and investigate a new augmentation framework for object detection datasets and train using faster-rcnn with ZF network as a backbone. We measure the Average Precision (AP) as stated in paper [PASCAL VOC] and show the correlation matrix for each class. Our findings show that data augmentation improving the performance for true positive. However, many false positive also occur but decreased by 19.7% after augmentation.