Receptive Field Block Net for Accurate and Fast Object Detection

last modified : 09-11-2018

General Information

  • Title: Receptive Field Block Net for Accurate and Fast Object Detection
  • Authors: Songtao Liu, Di Huang, Yunhong Wang
  • Link: article
  • Date of first submission: Tue, 21 Nov 2017
  • Implementations:

Brief

This article tries to improve the accuracy of the single shot detectors while avoiding a decrease of the speed of inference. To do so, they introduce a new module, the RF Block (RFB) inspired by the Receptive Fields (RFs) in the human visual systems. The idea is to copy the pattern of the neurones in the human visual cortex. They get results on par with the state of the art multi-stage detectors while keeping the speed advantage of the one stage classifiers.

How Does It Work

The architecture of the proposed test network is shown in the following figure and is a SSD with some of the layers replaced by the RFB module. The backbone network is still a VGG16 and the prediction is still made by a cascade of convolutions.

RFB Network

The idea is to look wider, the RBF module makes use of a trou convolutions to increase the perception field of the network, as shown in the following figure:

RFB module activation

Results

Comparison of results for detection on the VOC 2007:

Model mAP fps
Faster RCNN 73.2% 7
R-FCN 80.5% 9
YOLOv2 544 78.6% 40
R-FCN w deformable CNN 82.6 \% 8
SSD300 77.2% 46
RFB Net300 80.5% 43
RFB Net512 82.2% 18

Comparison of results for classification on COCO dataset:

Method Backbone Data Time Avg. Precision, IoU: 0.5:0.9
Faster VGG trainval 147 ms 24.2
Faster+++ ResNet-101 trainval 3.36 s 34.9
Faster w FPN ResNet-101-FPN trainval35k 240 ms 36.2
Faster by G-RMI Inception-Resnet-v2 trainval – 34.7
R-FCN ResNet-101 trainval 110 ms 29.9
R-FCN w Deformable CNN ResNet-101 trainval 125 ms 34.5
Mask R-CNN ResNext-101-FPN trainval35k 210 ms 37.1
YOLOv2 darknet trainval35k 25 ms 21.6
SSD300* VGG trainval35k 22 ms 25.1
SSD512* VGG trainval35k 53 ms 28.8
DSSD513 ResNet-101 trainval35k 182 ms 33.2
RetinaNet500 ResNet-101-FPN trainval35k 90 ms 34.4
RetinaNet800 ResNet-101-FPN trainval35k 198 ms 39.1
RFB Net300 VGG trainval35k 25 ms 29.9
RFB Net512 VGG trainval35k 55 ms 33.8
RFB Net512-E VGG trainval35k 59 ms 34.4

In Depth

They introduce two different module, the rfb_a and rfb_b. In conception they are both very similar, the internal layer is the only part changing, cf figures.

RFB-a: RFB module a

RFB-b: RFB module activation

For each of the two modules, the principal change compared to an inception module is the presence of the dilated convolutions.