Asymmetric Encoder-Decoder Structured FCN Based LiDAR to Color Image Generation.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Light Detection and Ranging (LiDAR) provides 3D structural information but lacks color.
• Generating color images from sparse LiDAR data is challenging.
• Existing methods like interpolation and Generative Adversarial Networks (GANs) have limitations.

Purpose of the Study:

• To propose a novel method for generating color images from LiDAR 3D reflection intensity.
• To develop a Fully Convolutional Network (FCN) capable of handling sparse input data.
• To evaluate the performance of the proposed method against conventional techniques.

Main Methods:

• A two-step approach: projecting LiDAR 3D reflection intensity to 2D, then using an FCN for color image generation.
• Designing an FCN with an asymmetric structure (deeper decoder than encoder) to manage sparse data.
• Training and evaluating the FCN on the KITTI dataset.

Main Results:

• The proposed method generates color images with good visual quality and accurate color fidelity.
• The FCN demonstrates superior performance compared to interpolation methods and Pix2Pix.
• Generated images are notably shadow-free and appear as daylight images.

Conclusions:

• The asymmetric FCN effectively generates color images from sparse LiDAR intensity data.
• The method offers a significant improvement over existing techniques for LiDAR-based colorization.
• The inherent properties of LiDAR data enable the generation of shadow-free, consistent color images.