티스토리 뷰

[업데이트 2018.09.05 14:47]


1. 베이스 논문

[논문 요약17] Perceptual Generative Adversarial Networks for Small Object Detection

(http://arclab.tistory.com/172)


Some efforts [4, 25, 18, 39, 23, 1] have been devoted to addressing small object detection problems. 

Some others [39, 23, 1] focus on developing network variants to generate multi-scale representation which enhances high-level small-scale features with multiple lower-level features layers.


2. 레퍼런스 논문

[1] S. Bell, C. L. Zitnick, K. Bala, and R. Girshick. Inside-outside net: Detecting objects in context with skip pooling and recurrent neural networks. arXiv preprint arXiv:1512.04143, 2015. 1, 6

(https://arxiv.org/pdf/1512.04143.pdf)


3. 주요 내용 요약

3.1 레퍼런스 논문 주요 내용

- 본 논문의 ION 아키텍쳐는 Multi-scale skip pooling과 context 정보(IRNN)를 사용하여 object detection을 수행합니다.


. Multi-scale representation

여러개의 lower-level convolutional layers로부터 ROI pooling을 통해 feature map을 추출하는 것을 의미합니다.


. Skip-layer connections

아래 그림은 본 논문에서 제안한 Inside-Outside Net 아키텍쳐입니다. 여러개의 lower-level convolutional layers와 4-directional IRNN(context feature)간에 skip-layer connection 개념(low-level feature와 high-level feature를 연결)에 따라 각각의 feature를 연결하며, 다음과 같은 과정을 거쳐 최종적으로 object detection 및 bounding box regression을 수행하게 됩니다. 


"L2 normalized, concatenated, scaled, dimension-reduced(1x1 convolution), fc, softmax, bounding box"



* Multi-scale representation 인용 논문

J. J. Koenderink and A. J. van Doorn. Representation of local geometry in the visual system. Bio. cybernetics, 1987.

(https://link.springer.com/article/10.1007/BF00318371)


3.2 베이스 논문에서 인용한 내용

- Small object detection의 성능 향상을 시도하였으나, Perceptual Generative Adversarial Networks for Small Object Detection 논문에서는 본 논문의 방법은 data augmentation이나 간단히 feature의 차원을 늘리는 방법으로 성능 향상을 시도하여 computational cost가 training/testing시 증가하는 단점이 있다고 이야기 하고 있습니다. 또한 low-level feature로부터 multi-scale representation을 통한 성능 향상 방법은 object detection시 한계가 있다고 말하고 있습니다.


- 아래와 같이 유사한 방법(multi-scale representation + skip pooling)이 사용되었습니다.



댓글
공지사항
최근에 올라온 글
최근에 달린 댓글
Total
Today
Yesterday
링크
«   2024/05   »
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 31
글 보관함