The Effect of Unfine-Tuned Super-Resolution Networks Act on Object Detection

In order to explore approaches for improving object detection accuracy in intelligent vehicle system, we exploit super-resolution techniques. A novel method is proposed to confirm the conjecture whether some popular super-resolution networks used for environmental perception of intelligent vehicles and robots can indeed improve the detection accuracy. COCO dataset which contains images from complex ordinary environment is utilized for the verification experiment, due to it can adequately verify the generalization of each algorithm and the consistency of experimental results. Using two representative object detection networks to produce the detection results, namely Faster R-CNN and YOLOv3, we devise to reduce the impact of resizing operation. The two networks allow us to compare the performance of object detection between using original and super-resolved images. We quantify the effect of each super-resolution techniques as well. Shown from our experimental results, the super-resolution networks provide average 3%-12% decrease in mAP (mean Average Precision) on enhancement level of ×2 and average 8%-19% decrease on enlargement level of ×4 compared to the conventional methods.