RefineDet is a novel, single-shot-based object detection algorithm. It achieves better accuracy than two-stage methods such as R-CNN and R-FCN but maintains comparable efficiency of one-stage method such as SSD and YOLOv3. For more information on RefineDet, see this paper:
RefineDet consists of two interconnected modules - the anchor refinement module and the object detection module. This improves the architecture of the one-stage approach to overcome the class imbalance problem and improve detection accuracy.
According to the paper, with the Pascal VOC dataset (a mainstream benchmark in visual object recognition and detection, when using a smaller input size of 320 x 320, RefineDet produces 80% mAP (mean Average Precision, a popular metric to measure the accuracy of object detectors).
By using a larger input size of 512x512, RefineDet achieves 81.8% mAP. This demonstrates that RefineDet has a better accuracy than other popular one-stage models such as SSD, Yolo for small-size object detection.
We have successfully deployed the RefineDet model on our Xilinx ZCU102 platform with the following performance and mAP result -
Platform: ZU9EG with 3x B4096 DPU
RefineDet computation: 25GOPS per image (480 x 360), unpruned
mAP (person class AP) on COCO sub2000 dataset: 63.39%Pre-processing time:3ms
DPU processing: 17ms
Post-processing time: 6.5ms
Throughput: 100fps with 8 threads
The RefineDet example is available in the DNNDK v3.0 release.