eduzhai > Applied Sciences > Engineering >

CoDeNet Efficient Deployment of Input-Adaptive Object Detection on Embedded FPGAs

  • Save

... pages left unread,continue reading

Document pages: 11 pages

Abstract: Deploying deep learning models on embedded systems has been challenging dueto limited computing resources. The majority of existing work focuses onaccelerating image classification, while other fundamental vision problems,such as object detection, have not been adequately addressed. Compared withimage classification, detection problems are more sensitive to the spatialvariance of objects, and therefore, require specialized convolutions toaggregate spatial information. To address this need, recent work introducesdynamic deformable convolution to augment regular convolutions. However, thiswill lead to inefficient memory accesses of inputs with existing hardware. Inthis work, we harness the flexibility of FPGAs to develop a novel objectdetection pipeline with deformable convolutions. We show the speed-accuracytradeoffs for a set of algorithm modifications including irregular-accessversus limited-range and fixed-shape. We then Co-Design a Network CoDeNet withthe modified deformable convolution and quantize it to 4-bit weights and 8-bitactivations. With our high-efficiency implementation, our solution reaches 26.9frames per second with a tiny model size of 0.76 MB while achieving 61.7 AP50on the standard object detection dataset, Pascal VOC. With our higher accuracyimplementation, our model gets to 67.1 AP50 on Pascal VOC with only 2.9 MB ofparameters-20.9x smaller but 10 more accurate than Tiny-YOLO.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...