Skip to main content

Region-Based Convolutional Networks for Accurate Object Detection and Segmentation

Citations: 1,514 | Paper Published: 01 Jan 2016 | Institutions: Microsoft, University of California, Berkeley | Authors: Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra Malik

Abstract in 2 Lines

The paper proposes a simple and scalable object detection algorithm that combines high-capacity convolutional networks with bottom-up region proposals to achieve better performance on the PASCAL VOC object detection task. The proposed model is called R-CNN or Region-based Convolutional Network.

Contributions​

  • Proposing a simple and scalable object detection algorithm that combines high-capacity convolutional networks with bottom-up region proposals to achieve better performance on the PASCAL VOC object detection task.
  • Introducing the concept of supervised pre-training for an auxiliary task, followed by domain-specific fine-tuning, to boost performance significantly when labeled training data are scarce.
  • Providing a new state-of-the-art performance on the PASCAL VOC 2012 object detection task with a mean average precision (mAP) of 62.4 percent.

Practical Implications​

  • The proposed R-CNN model can be used for object detection tasks in various applications such as autonomous driving, surveillance, and robotics.
  • The supervised pre-training approach can be used to improve the performance of object detection models when labeled training data are scarce.
  • The proposed algorithm can be used as a baseline for future research in object detection and related fields.

Methods​

Methodology
  • The proposed R-CNN model combines high-capacity convolutional networks with bottom-up region proposals to achieve better performance on the PASCAL VOC object detection task.
  • The model is trained in two stages: (1) pre-training on a large dataset for an auxiliary task, and (2) fine-tuning on the target object detection dataset.
  • The selective search algorithm is used to generate region proposals for the object detection task.
  • The model is trained using a multi-task loss function that combines classification and bounding box regression losses.

Data​

The paper has used the PASCAL VOC 2012 object detection dataset to evaluate the proposed R-CNN model. Additionally, the paper has also presented results on the ILSVRC2013 detection dataset.

Results​

Performance Analysis
  • The proposed R-CNN model achieved a mean average precision (mAP) of 62.4%.
  • The R-CNN model outperformed four strong baselines, including SegDPM.
  • The paper also reports results on the ILSVRC2013 detection dataset, where the R-CNN model achieved a mAP of 31.4%.

Limitations​

  • The R-CNN model is computationally expensive and slow at test time, as it requires running a CNN on each region proposal. This limits its practical use in real-time applications.
  • The model requires a large amount of GPU memory during training, which can be a bottleneck for some users.

Table of Contents