This study evaluates road surface object detection tasks using four Mask R-CNN models available on the Tensor-Flow Object Detection API. The models were pre-trained using COCO datasets and fine-tuned by 15,1SS segmented road surface annotation tags. Validation data set was used to obtain Average Precisions and Average Recalls. Result indicates a substantial false negatives or "left judgement"counts for linear cracks, joints, fillings, potholes, stains, shadows and patching with grid cracks classes. There were significant number of incorrectly predicted label instances. To improve the result, an alternative metric calculation method was tested. However, the results showed strong mutual interferences caused by misinterpretation of the scratches with other object classes. © 2020 IEEE.