Exploiting Target Data to Learn Deep Convolutional Networks for Scene-Adapted Human Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Generic human detectors perform poorly in real-world scenes due to distribution differences.
• Manual annotation for scene-specific datasets is labor-intensive and time-consuming.

Purpose of the Study:

• To develop a semi-supervised approach for training deep convolutional networks on partially labeled data.
• To improve human detection performance in specific, real-world scenes.

Main Methods:

• Knowledge transfer from public datasets by adapting an auxiliary detector to the target scene.
• A selective ensemble algorithm to identify and recombine relevant detector components.
• Progressive labeling and self-paced sample selection for training deep convolutional networks.

Main Results:

• The proposed approach effectively adapts models to target scenes, enhancing detection accuracy.
• The selective ensemble method successfully identifies scene-relevant human characteristics.
• The semi-supervised method significantly outperforms generic detectors in scene-specific human detection tasks.

Conclusions:

• The developed semi-supervised method offers an effective solution for scene-specific human detection.
• This approach reduces reliance on extensive manual annotations by leveraging unlabeled data.
• The technique demonstrates superiority and effectiveness in real-world application scenarios.