Small Object Detection Using Depth based Optimized NMS
Ravi Prakash Chaturvedi, Udayan Ghose · 2023
This paper introduces the method and algorithm for improvement of detection system based on CNN. due to some poor resolution and conflict background system faces problem of small object detection. So this paper proposed hybrid method for small object detection. Using improved loss function on IOU for bounding box regression and solve the positioning deviations by using ROI pooling operation and also use the multi-scale convoluted fussed features for more information. This algorithm gives good performance on small object like knife, cutter, wallet, screwdriver etc. Algorithm recall rate reach 85% and accuracy reach 80%. So, the detection is significantly better than other method and more effective to detect small object.