Skip to main content

Deep Neural Networks for Object Detection

Citations: 1,172 | Paper Published: 05 Dec 2013 | Institutions: Google | Authors: Christian Szegedy, Alexander Toshev, Dumitru Erhan

Abstract in 2 Lines

The paper introduces a Region Proposal Network (RPN) that generates high-quality region proposals for object detection. The RPN shares full-image convolutional features with the detection network, enabling nearly cost-free region proposals and achieving state-of-the-art accuracy on various datasets.

Contributions​

  • Presenting a formulation for object detection using DNNs that can predict the bounding boxes of multiple objects in a given image.
  • Defining a DNN-based regression that outputs a binary mask of the object bounding box (and portions of the box as well).
  • Employing a simple bounding box inference to extract detections from the masks.
  • Applying the DNN mask generation in a multi-scale fashion on the full image as well as on a small number of large image crops, followed by a refinement step.
  • Showing state-of-the-art performance of the approach on Pascal VOC.

Practical Implications​

  • The proposed method can be used for object detection in various applications such as autonomous driving, surveillance, and robotics.
  • The method can detect multiple objects of different classes in a given image with high accuracy.
  • The multi-scale inference procedure reduces the computational cost of the method while maintaining high-resolution object detections.
  • The approach can be extended to detect objects of different classes using a single network, which can further reduce the computational cost.

Methods​

Methodology
  • Using a pre-trained Deep Neural Network (DNN) for classification as a starting point.
  • Replacing the last layer of the DNN with a regression layer to predict object bounding box masks.
  • Defining a multi-scale inference procedure to produce high-resolution object detections at a low cost by a few network applications.

Data​

The paper has used the Pascal VOC dataset for evaluating the proposed method. The Pascal VOC dataset is a widely used benchmark dataset for object detection, which contains 20 object classes and 11,530 images for training and validation.

Results​

Performance Analysis
  • Mean Average Precision (mAP): It measures the average precision of the model across all object classes and detection thresholds. The proposed method achieved a mAP of 74.3%, which is a significant improvement over the previous state-of-the-art method.
  • Intersection over Union (IoU): It measures the overlap between the predicted bounding box and the ground truth bounding box. The proposed method achieved an IoU of 55.7%, which is also a significant improvement over the previous state-of-the-art method.

Limitations​

  • The proposed method requires a large amount of training data to achieve high accuracy, which can be a limitation in some applications where labeled data is scarce.
  • The method is computationally expensive, especially during training, which can be a limitation for resource-constrained devices.

Table of Contents