[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-优快云博客

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-优快云博客

4. Reference

Liu, S. et al. (2018) Receptive Field Block Net for Accurate and Fast Object Detection, ECCV.