Fully Automated DCNN-Based Thermal Images Annotation Using Neural Network Pretrained on RGB Data

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep neural networks (DNNs) for object detection demand extensive annotated data, often millions of images.
• Large-scale RGB image datasets are abundant, but comparable thermal image datasets are scarce.
• This limitation hinders the training of state-of-the-art DNNs for thermal imaging applications.

Purpose of the Study:

• To present a novel, fully automatic method for generating large-scale annotated thermal image datasets.
• To enable the training of high-performance object detectors using thermal imagery.
• To address the bottleneck of manual data annotation in thermal domain.

Main Methods:

• Utilized an RGB pre-trained object detector in conjunction with RGB and thermal cameras, and 3D LiDAR.
• Developed a fully automated pipeline for thermal image labeling.
• Created a dataset with hundreds of thousands of annotated objects through this automated process.

Main Results:

• Successfully generated a large-scale annotated thermal image dataset.
• Trained deep learning models achieving performance comparable to human-annotation-based methods.
• Demonstrated significantly improved results compared to training with small-scale, hand-annotated thermal datasets.

Conclusions:

• The proposed automated method effectively overcomes the challenge of limited annotated thermal data.
• This approach facilitates the development of more robust and accurate object detection models for thermal imaging.
• The generated dataset and methodology offer a valuable resource for advancing research in thermal computer vision.