Highlights

  • Network generates region-orientated attention maps and creates an informative translation-variant representation of RoI features.
  • Simplified subnetwork to a single low-capacity FC layer, which largely reduces parameter size and speeds up inference.
  • Improved detection performance using ResNet-101, GoogLeNet and VGG-16code backbones.

Summary

The goal is to extract effective RoI features from translation-invariant convolutional feature maps. A two-stage object detection framework that consists of a proposal generator and a region classifier is used.

The paper proposes feature selective networks, which introduce dimension reduction and region-wise feature attention. RoI features are extracted with respect to the sub-region variation and aspect ratio preference.

Introduction

Most of state-of-the-art object detectors adopt two-stage framework composed of a proposal generator and a region classifier. RoI features are extracted by generating whole image feature maps by ConvNets, and then pooling the RoI features.

Classical RoI pooling layer divides RoIs into pooling bins and max-pools. RoI features originating from convolutional feature map are high-dimensional.

RoI features pooled by classical RoI pooling are translation-invariant, i.e. insensitive to detection tasks. RoI features are extracted independently for different sub-regions and aspect ratio.

Methods

  • Generate feature map using ConvNet (ImageNet pre-trained classification backbones).
  • Channel reduction using 1x1 convolution.
  • Max-pool.
  • Compute RoIs provided by Region Proposal Network (RPN).
  • Generate sub-region attention maps using shifted convolutional layers on the convolutional feature map of the entire image. For the same spatial position inside different sub-region, the sub-region attention map gives different feature attentions.

  • Generate aspect ratio attention map for each RoI using a 1×1 convolutional layer on the convolutional feature map. For the RoIs of different aspect ratios, the aspect ratio attention map also produces distinct feature attentions.

  • Selective RoI pooling.
  • Detection subnetwork.

Data

  • PASCAL VOC 2007, VOC 2012
  • MS COCO val2015

Benchmarking

  • Faster R-CNN
  • R-FCN

Results