The only framework which can provide 45 FPS (on<br>GPU) and mAP value of about 63.4% on VOC2007 (real-time<br>data) was YOLO. Still it faces problems in identifying of<br>smaller objects in that frame. This problem was rectified using<br>SSD [8] which follows the policy of combining anchor box<br>proposal system of faster-RCNN and uses muti-scale features<br>for performing detection layer. The mAP value on VOC2007<br>was increased to 73.9% by preserving the detection speed<br>same as that of YOLO.
การแปล กรุณารอสักครู่..
