Object Detection Networks on Convolutional Feature Maps.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Object detectors typically comprise feature extractors and object classifiers.
• Feature extractors have advanced significantly with deep convolutional architectures.
• Object classifiers, often simple multi-layer perceptrons, have lagged in development.

Purpose of the Study:

• To highlight the importance of designing deep networks for object classification.
• To introduce and evaluate Networks on Convolutional feature maps (NoCs) for object detection.

Main Methods:

• Experimenting with region-wise classifier networks utilizing shared, region-independent convolutional features (NoCs).
• Comparing the performance of NoCs against existing state-of-the-art image classification models in object detection tasks.

Main Results:

• Deep convolutional feature maps are essential, but a deep, convolutional per-region classifier is critically important.
• Advanced image classification models (e.g., ResNets, GoogLeNets) do not inherently improve detection accuracy without a specialized per-region classifier.
• NoC design proved essential for top-performing entries in the 2015 ImageNet and MS COCO challenges.

Conclusions:

• Careful design of deep networks for object classification is as vital as feature extraction.
• Networks on Convolutional feature maps (NoCs) represent a key architectural component for state-of-the-art object detection.