[ECCV 2018]Receptive Field Block Net for Accurate and Fast Object Detection
论文网址：[1711.07767] Receptive Field Block Net for Accurate and Fast Object Detection
论文代码：GitHub - GOATmessi8/RFBNet: Receptive Field Block Net for Accurate and Fast Object Detection, ECCV 2018
英文是纯手打的！论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误，若有发现欢迎评论指正！文章偏向于笔记，谨慎食用
目录
1. 心得
2. 论文逐段精读
2.1. Abstract
2.2. Introduction
2.3. Related Work
2.4. Method
2.4.1. Visual Cortex Revisit
2.4.2. Receptive Field Block
2.4.3. RFB Net Detection Architecture
2.4.4. Training Settings
2.5. Experiments
2.5.1. Pascal VOC 2007
2.5.2. Ablation Study
2.5.3. Microsoft COCO
2.6. Discussion
2.7. Conclusion
3. 知识补充
3.1. Atrous Spatial Pyramid Pooling（ASPP）
3.2. Deformable Convolution
4. Reference
1. 心得
（1）比较简单易懂的模块
2. 论文逐段精读
2.1. Abstract
①Challenges: deep CNNs have higher accuracy but run slowly, lightweight models are often bad in performance
②Solving methods: so they design a feature enhancement method on lightweight model to ensure accuracy
eccentricity n.古怪；怪癖；反常；古怪行为；[数]离心率
2.2. Introduction
①The size of population Receptive Field (pRF) of human is eccentricity in retinotopic maps:
②They proposed a lightweight Receptive Field Block (RFB), and assemble it to the top of SSD to get a one-stage detector (RFB Net):
they simulate eccentricities by different dilated rate
2.3. Related Work
①Lists two stage and one stage models
②Difference of typical RF models, Inception, ASPP, Deformable Conv and RFB:
2.4. Method
2.4.1. Visual Cortex Revisit
①Combine fMRI and pRF, researchers can find the correlation between cortex and visual field maps
②There is a positive relationship between pRF and eccentricity
2.4.2. Receptive Field Block
①Design of RFB:
2.4.3. RFB Net Detection Architecture
①Pipeline of RFB-Net300:
2.4.4. Training Settings
①在下一节说明
2.5. Experiments
①Datasets: Pascal VOC 2007 and MS COCO
②Categories: 20 and 80
2.5.1. Pascal VOC 2007
①Batch size: 32
②Initial learning rate: 1e-3, warming up from 1e-6 to 4e-3 at the first 5 epochs
③Epoch: 250
④Weight decay: 0.0005
⑤Momentum: 0.9
⑥Comparison table:
2.5.2. Ablation Study
①Module ablation study:
②Comparison with other architectures:
2.5.3. Microsoft COCO
①Set: trainval35k set (train set + val 35k set)
②Batch size: 32
③Warm up: from 1e-6 to 2e-3 in the first 5 epoch and reduce it after 80 and 100 epochs by the factor of 10, and end up at 120.
④Performance table:
②Module ablation on MobileNet:
2.6. Discussion
①Inference time:
2.7. Conclusion
~
3. 知识补充
3.1. Atrous Spatial Pyramid Pooling（ASPP）
（1）参考学习：ASPP 详解-CSDN博客
3.2. Deformable Convolution
（1）参考学习：CNN卷积神经网络之DCN（Deformable Convolutional Networks、Deformable ConvNets v2）_dcn神经网络-CSDN博客
4. Reference
Liu, S. et al. (2018) Receptive Field Block Net for Accurate and Fast Object Detection, ECCV.