티스토리 뷰

[업데이트 2019.08.24 14:19]

 

1. 논문

M2det: A single-shot object detector based on multi-level feature pyramid network

 

Zhao, Q., Sheng, T., Wang, Y., Tang, Z., Chen, Y., Cai, L., & Ling, H. (2019). M2Det: A Single-Shot Object Detector Based on Multi-Level Feature Pyramid Network. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01), 9259-9266. https://doi.org/10.1609/aaai.v33i01.33019259

Published 2019-07-17

 

2. 요약

Feature pyramids are widely exploited by both the state-ofthe-art one-stage object detectors (e.g., DSSD, RetinaNet, RefineDet) and the two-stage object detectors (e.g., Mask RCNN, DetNet) to alleviate the problem arising from scale variation across object instances.

- Newly, in this work, we present Multi-Level Feature Pyramid Network (MLFPN) to construct more effective feature pyramids for detecting objects of different scales.

First, we fuse multi-level features (i.e. multiple layers) extracted by backbone as the base feature. 

Second, we feed the base feature into a block of alternating joint Thinned U-shape Modules and Feature Fusion Modules and exploit the decoder layers of each Ushape module as the features for detecting objects. Finally, we gather up the decoder layers with equivalent scales (sizes) to construct a feature pyramid for object detection, in which every feature map consists of the layers (features) from multiple levels.

 

 

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