Scribd
Upload
21 views

Object Detection Over FPGA

Object detection is a technique used to identify and locate objects in images or videos. It can be useful for counting objects, determining their precise locations, and tracking them while accurately labeling them. Recent models like YOLOv3 provide high detection precision with low inference time, making it suitable for real-time object detection applications. Implementing object detection models like YOLO on FPGAs can provide higher frame rates compared to CPU implementations, important for applications requiring minimum 30 FPS such as live video analysis. However, smaller models like YOLOv3-tiny provide a good balance of accuracy and efficiency for resource-constrained hardware.

Uploaded by

rahul r
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
21 views

Object Detection Over FPGA

Object detection is a technique used to identify and locate objects in images or videos. It can be useful for counting objects, determining their precise locations, and tracking them while accurately labeling them. Recent models like YOLOv3 provide high detection precision with low inference time, making it suitable for real-time object detection applications. Implementing object detection models like YOLO on FPGAs can provide higher frame rates compared to CPU implementations, important for applications requiring minimum 30 FPS such as live video analysis. However, smaller models like YOLOv3-tiny provide a good balance of accuracy and efficiency for resource-constrained hardware.

Uploaded by

rahul r
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Object

Detection over
FPGA
-Presented by Amit Chougule ([email protected])

This Photo by Unknown author is licensed under CC BY-SA.


• It is a technique that allows us to identify and locate objects in an image or
Object video.

Detection • object detection can be useful for count objects in a scene and determine
and track their precise locations, all while accurately labeling them.
Year Model
Available Object
2014 OverFeat, R-CNN.
Detection models
2015 Fast R-CNN ,Faster R-CNN, MR-CNN & S-
CNN, and DeepID-Net.
2016 CRAFT, R-FCN, ION, MultiPathNet,
Hikvision, GBD-Net / GBD-v1 & GBD-v2,
SSD, YOLOv1.
2017 NoC, G-RMI, TDM, DSSD, YOLOv2 /
YOLO9000, FPN, RetinaNet, DCN / DCNv1,
Light-Head R-CNN.
2018 YOLOv3, Cascade R-
CNN, MegDet, StairNet.
2019 DCNv2.
Comparison
of models
*mAP = mean average precision

• FPN FRCN gives high precision (59.1), but also its inference
time is high, which may makes object detection difficult for
the real-time scenario.
• In 2nd place, YOLOv3 gives higher precision(57.9) with
lowest inference time, which is perfectly suitable for real-
time object detection.
Versions of YOLO

Computer Vision
library
Frames Per Second
(FPS)
• FPS is used to measure frame rate – the
number of consecutive full-screen images
that are displayed each second.
• In the case of Object detection, if fps is less,
the model will predict the object's location &
label in slower late, which become problem
for decision taking system in autonomous
driving car.
Recent Work
Paper FPGA FPS Link
1. YOLO Acceleration using FPGA Zynq board (which is 19 https://ieeexplore.ieee.org/stamp/s
Architecture (IEEE) released by Xillinx) tamp.jsp?arnumber=8641256
2. Sparse-YOLO: Hardware/Software Co- - - https://ieeexplore.ieee.org/stamp/s
Design of an FPGA Accelerator for tamp.jsp?arnumber=9122495
YOLOv2 (IEEE)
CPU+ FPGA
3. REQ-YOLO: A Resource-Aware, Efficient - - https://arxiv.org/pdf/1909.13396.pd
Quantization Framework for Object f
Detection on FPGAs
4. A Demonstration of FPGA-based You Xilinx Inc. zcu102 35 http://kalman.mee.tcd.ie/fpl2018/c
Only Look Once version2 (YOLOv2) board, ontent/pdfs/FPL2018-
(which has the Xilinx 43iDzVTplcpussvbfIaaHz/PdMR0d2Bf
Inc. Zynq Ultrascale+ gpX7hSugYkVV/5FYmmZ6wsv324TEr
MPSoC) srdsBw.pdf
• For Live camera or streaming minimum 30 fps is required.
So, while hardware selection or object detection model
development time, we need to maintain 30+ fps.
• YOLOv3 provides, 30 fps over Pascal Titan X, whereas
YOLOv4 provides 65 FPS on a Tesla V100.
• Also, for hardware - cost, memory & power reduction
purpose, we can use tiny version of either YOLOv3-tiny or
Key-points YOLOv4-tiny.
• "Tiny" is a simplified version of YOLO, which has a much
smaller number of convolution layers than YOLO.
• That means that tiny-yolo does not need to occupy a large
amount of memory, reducing the need for hardware. And it
also greatly speeds up detection, but lost some of the
detection accuracy.
Thank you

You might also like