The comparison of Faster R-CNN and Atrous Faster R-CNN in different distance and light condition

Abstract

This paper presents the comparison of Faster R-CNN and Atrous Faster R-CNN, which detection model, in the different distance and light condition. Also, the dataset for model training is COCO, and the classification model is residual network. The parameter for decision the performance of the model is Mean Average Precision (mAP). The results from an object resolution at 1024x768 of Faster R-CNN at 3 meters in the evening achieved mAP 1.000. Besides, the mAP at 5 meters and 8 meters were 0.798 and 0.760, respectively. The same resolution as previous, the results of Atrous Faster R-CNN at 3 meters in the evening presented mAP 1.000. Also, the mAP at 5 meters and 8 meters were 1.000 and 0.960, respectively. In addition, Atrous Faster R-CNN had better accuracy than Faster R-CNN with appropriate range and brightness from the period of the day for real-life usage.