Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Object Detection Networks on Convolutional Feature Maps

Shaoqing Ren, Kaiming He, Ross Girshick, Xiangyu Zhang, Jian Sun

TL;DR

This work introduces Networks on Convolutional feature maps (NoCs), a class of region-wise classifiers built on RoI-pooled convolutional features, and demonstrates that a deep, convolutional NoC is essential for high object-detection accuracy. Through extensive ablations on VOC07 with ZF and VGG backbones, it shows that deeper NoCs and conv-based classifiers outperform traditional MLPs and that scale-aware maxout further improves performance. Combining NoCs with Faster R-CNN architectures, including ResNet and GoogLeNet backbones, yields substantial gains on MS COCO and VOC benchmarks, arguing that classifier design on RoI-pooled features is as important as feature extraction depth. The findings advocate integrating deep, convolutional NoCs into modern detectors to achieve top performance on challenging datasets.

Abstract

Most object detectors contain two important components: a feature extractor and an object classifier. The feature extractor has rapidly evolved with significant research efforts leading to better deep convolutional architectures. The object classifier, however, has not received much attention and many recent systems (like SPPnet and Fast/Faster R-CNN) use simple multi-layer perceptrons. This paper demonstrates that carefully designing deep networks for object classification is just as important. We experiment with region-wise classifier networks that use shared, region-independent convolutional features. We call them "Networks on Convolutional feature maps" (NoCs). We discover that aside from deep feature maps, a deep and convolutional per-region classifier is of particular importance for object detection, whereas latest superior image classification models (such as ResNets and GoogLeNets) do not directly lead to good detection accuracy without using such a per-region classifier. We show by experiments that despite the effective ResNets and Faster R-CNN systems, the design of NoCs is an essential element for the 1st-place winning entries in ImageNet and MS COCO challenges 2015.

Object Detection Networks on Convolutional Feature Maps

TL;DR

This work introduces Networks on Convolutional feature maps (NoCs), a class of region-wise classifiers built on RoI-pooled convolutional features, and demonstrates that a deep, convolutional NoC is essential for high object-detection accuracy. Through extensive ablations on VOC07 with ZF and VGG backbones, it shows that deeper NoCs and conv-based classifiers outperform traditional MLPs and that scale-aware maxout further improves performance. Combining NoCs with Faster R-CNN architectures, including ResNet and GoogLeNet backbones, yields substantial gains on MS COCO and VOC benchmarks, arguing that classifier design on RoI-pooled features is as important as feature extraction depth. The findings advocate integrating deep, convolutional NoCs into modern detectors to achieve top performance on challenging datasets.

Abstract

Most object detectors contain two important components: a feature extractor and an object classifier. The feature extractor has rapidly evolved with significant research efforts leading to better deep convolutional architectures. The object classifier, however, has not received much attention and many recent systems (like SPPnet and Fast/Faster R-CNN) use simple multi-layer perceptrons. This paper demonstrates that carefully designing deep networks for object classification is just as important. We experiment with region-wise classifier networks that use shared, region-independent convolutional features. We call them "Networks on Convolutional feature maps" (NoCs). We discover that aside from deep feature maps, a deep and convolutional per-region classifier is of particular importance for object detection, whereas latest superior image classification models (such as ResNets and GoogLeNets) do not directly lead to good detection accuracy without using such a per-region classifier. We show by experiments that despite the effective ResNets and Faster R-CNN systems, the design of NoCs is an essential element for the 1st-place winning entries in ImageNet and MS COCO challenges 2015.

Paper Structure

This paper contains 13 sections, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Ablation Experiments
  4. Using MLP as NoC
  5. Using ConvNet as NoC
  6. Maxout for Scale Selection
  7. Fine-tuning NoC
  8. Deep Features vs. Deep Classifiers
  9. Error Analysis
  10. Comparisons of Results
  11. Summary of Observations
  12. NoC for Faster R-CNN with ResNet
  13. Conclusion

Figures (3)

  • Figure 1: Overview of NoC. The convolutional feature maps are generated by the shared convolutional layers. A feature map region is extracted and RoI-pooled into a fixed-resolution feature. A new network, called a NoC, is then designed and trained on these features. In this illustration, the NoC architecture consists of two convolutional layers and three fully-connected layers.
  • Figure 2: A maxout NoC of "c256-mo-c256-f4096-f4096-f21". The features are RoI-pooled from two feature maps computed at two scales. In this figure, maxout is used after conv$_{+1}$.
  • Figure 3: Distribution of top-ranked True Positives (TP) and False Positives (FP), generated by the published diagnosis code of Hoiem2012. The types of positive predictions are categorized Hoiem2012 as Cor (correct), Loc (false due to poor localization), Sim (confusion with a similar category), Oth (confusion with a dissimilar category), BG (fired on background). The total number of samples in each disk is the same and equal to the total number of ground-truth labels Hoiem2012. More explanations are in the main text.