티스토리 뷰
[2019][***]An efficient convolutional neural network for small traffic sign detection
Arc Lab. 2019. 8. 1. 15:071. 인용 논문
Traffic-sign detection and classification in the wild
Zhe Zhu, Dun Liang, Songhai Zhang, Xiaolei Huang, Baoli Li, Shimin Hu;
The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2110-2118
2. 인용 부분
4.1. Architectural design strategies
Our overarching purpose is to build a CNN model with few model pa- rameters but a competitive detection accuracy. To this end, we employ following three improved strategies.
Strategy I. Factorize 5 × 5 filters into two 3 × 3 filters.
Given a budget of an exact number of convolution kernels, we select to deploy the majority of these 1×1 and 3×3 kernels, since a 3×3 kernel has 25 / 9 fewer parameters than a 5 × 5 kernel.
Strategy 2. Decrease the superfluous 3 × 3 convolutional layers and substitute 3 × 3 kernels with 1 × 1 kernels.
Following [33], we prune the last 3 × 3 convolutional layer after ex- periments, and then replace the remaining two 3 × 3 convolutional lay- ers with 1 × 1 filters. Although our feature extraction network includes 5 convolutional layers, whose number is the same as VGGNet [14] and model in [33], introduction of small kernels can remarkably reduce the parameters in the model because a 1 × 1 kernel has 9x fewer parameters than a 3 × 3 kernel.
Strategy 3. Transform the fully connected layers to 1×1 fully convolutional layers.
In a convolutional neutral network, fully connected layers usually function to output the confidence for each possible category, which con- sumes large amount of calculation cost and introduce numerous param- eters. Our modification is that the standard fully connected layers can be equally replaced by the 1×1 fully convolutional layers, which are more efficient and beneficial for decreasing computing consumption.
Strategies 1 and 2 are about advisably reducing the quantity of pa- rameters in a CNN while maintaining a competitive accuracy. Strategy 3 is about minimizing calculation amount on a limited budget of network. Next, we describe our small object detector for traffic signs, which is backboned on CNN architecture that activates us to successfully utilize Strategies 1, 2 and 3.
- Total
- Today
- Yesterday
- #ApacheZeppelin
- Ragdoll
- #REST API
- Mask R-CNN
- OST
- some time ago
- Worry
- Game Engine
- #TensorFlow
- #ApacheSpark
- Badge
- ate
- ILoop Engine
- aws #cloudfront
- #ELK Stack
- GOD
- docker
- Meow
- Physical Simulation
- 도커
- belief
- Library
- Jekyll and Hyde
- English
- 2D Game
- sentence test
- SSM
- project
- Sea Bottom
- Memorize
일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | 2 | |||||
3 | 4 | 5 | 6 | 7 | 8 | 9 |
10 | 11 | 12 | 13 | 14 | 15 | 16 |
17 | 18 | 19 | 20 | 21 | 22 | 23 |
24 | 25 | 26 | 27 | 28 | 29 | 30 |