School:National Tsing Hua University
Department: Computer Science
Student:張芸綺 Yun-Chi Chang
Advisor:李濬屹 Chun-Yi Lee
input image
prediction
bus?
car?
background?
Input image
Ground truth
DeepLab-VGGNet
DeepLab-ResNet-101
Input image
Ground truth
DeepLab-VGGNet
DeepLab-ResNet-101
Is input image well-performed?
Yes
No
Edge-end devices
Server
Small network on edge-end devices
Large network on remote server
Image recognition
Object detection
Bounding Box
Semantic Segmentation
RCNN Fast-RCNN
Union of the prediction and the ground truth
Overlap of the prediction and the ground truth
: Number of objects in the data set.
Union of the prediction and the ground truth
Overlap of the prediction and the ground truth
: Number of objects in the data set.
DeepLab-v3 add an image pooling layer in ASPP.
The architecture of our work contains four parts:
DeepLab-VGGNet (mIoU=69%)
DeepLab-ResNet-101 (mIoU=81%)
Our Goal
DeepLab-VGGNet
DeepLab-ResNet-101
The Decision Network contains:
(feature extraction)
The Decision Network contains:
(feature extraction)
The mIoU is defined as:
We define IoU_img to estimate our accuracy on an image:
: Number of classes contained in an image
: Number of objects in the data set.
0 0 0.95 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0.95 0 0 0.98 0 0 0 0 0 0 0 0 0 0 0 0 0 0
IoU_img = 0.95 / 1 = 0.95
IoU_img = (0.95 + 0.98) / 2
= 0.965
Image number
Image number
Image number
Image number
CU
To evaluate our decision network, we calculate :
true positive
false positive
To evaluate our decision network, we calculate :
true positive/ total images in the subset of images with IoU_img .
true negative/ total images in the subset of images with IoU_img .
Overall computation reduction (Decision-Network-AlexNet)
Edge-end computation reduction (Decision-Network-AlexNet)
Decision-Network-AlexNet
Decision-Network-ResNet-101