Light-Weight Object Detection Networks for Embedded Platform

Jiahua Cui; Yunzhou Zhang; Zheng Wang; Jiwei Liu

doi:10.3788/AOS201939.0415006

Journals >Acta Optica Sinica >Volume 39 >Issue 4 >Page 0415006 > Article

Acta Optica Sinica
Vol. 39, Issue 4, 0415006 (2019)

Light-Weight Object Detection Networks for Embedded Platform

Jiahua Cui¹, Yunzhou Zhang^1、2、*, Zheng Wang¹, and Jiwei Liu¹

Author Affiliations

¹ College of Information Science & Engineering, Northeastern University, Shenyang, Liaoning 110819, China;

² Faculty of Robot Science and Engineering, Northeastern University, Shenyang, Liaoning 110819, China

show less

DOI: 10.3788/AOS201939.0415006 Cite this Article Set citation alerts

Jiahua Cui, Yunzhou Zhang, Zheng Wang, Jiwei Liu. Light-Weight Object Detection Networks for Embedded Platform[J]. Acta Optica Sinica, 2019, 39(4): 0415006 Copy Citation Text

show less

Fig. 1. Flow chart of MTYOLO object detection algorithm

Download full size

Basic structure of depth separable convolution network. (a) Standard convolution filter; (b) depthwise convolution filter; (c) pointwise convolution filter

Fig. 2. Basic structure of depth separable convolution network. (a) Standard convolution filter; (b) depthwise convolution filter; (c) pointwise convolution filter

Download full size

Fig. 3. Architecture of MTYOLO network

Download full size

Fig. 4. Comparison of proposed algorithm and Tiny-Yolo algorithm. (a) Input image; (b) Tiny-Yolo algorithm; (c) proposed algorithm

Download full size

Conv2	Conv3	Conv4	Conv5	Conv6	Conv7	Conv10	P_mAP
×	×	×	×	×	×	×	54.2
×	×	×	√	√	×	×	56.7
×	×	√	√	√	×	×	57.6
×	√	√	√	√	×	×	58.4
√	√	√	√	√	×	×	58.2
×	√	√	√	√	√	×	58.9
×	√	√	√	√	√	√	59.3

Table 1. Comparison of detection accuracy after merging different layers

BN layer	P_mAP	Speed /(frame·s^-1)
×	52.7	31
√	56.2	29

Table 2. Comparison of experimental results of MTYOLO with BN layer

Model	Speed /(frame·s^-1)	Model size /MB
Yolo-v2	3	193
Tiny-Yolo	18	61
Using feature map fusion	15	63
Using pointwise	23	50
Using depthwise	29	41

Table 3. Comparison of experimental results of different network architectures

Method	P_mAP	FPS	aero	bike	bird	boat	bottle	bus	car	cat	chair
Tiny-Yolo	54.2	18	57.4	67.5	44.9	34.8	20.4	67.5	62.9	67.4	32.0
MTYOLO	57.25	29	65.3	74.2	49.6	39.1	30.2	69.8	65.8	69.5	34.2
Method	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	TV
Tiny-Yolo	53.7	58.1	61.6	70.5	69.1	58.0	27.8	52.8	51.1	68.5	57.4
MTYOLO	56.2	59.3	63.5	72.1	70.2	59.2	29.8	55.1	53.6	69.1	59.2

Table 4. Comparison of detection results on VOC dataset by proposed algorithm and Tiny-Yolo

Jiahua Cui, Yunzhou Zhang, Zheng Wang, Jiwei Liu. Light-Weight Object Detection Networks for Embedded Platform[J]. Acta Optica Sinica, 2019, 39(4): 0415006

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information