A self-supervised deep learning model for extrinsic calibration between LiDAR and camera sensors using 3D spatial transformer networks.
CalibNet is a self-supervised deep learning model for extrinsic calibration between LiDAR and camera sensors in autonomous systems. It estimates the 6-degree-of-freedom rigid transformation that aligns LiDAR point clouds with camera images without requiring manually labeled calibration data. The model uses 3D spatial transformer networks to learn this alignment through a point cloud distance loss function.
Researchers and engineers working on autonomous vehicles, robotics, or multi-sensor perception systems who need accurate and automated calibration between LiDAR and cameras.
It eliminates the need for tedious manual calibration or labeled datasets by using a self-supervised approach, provides an end-to-end trainable deep learning solution, and integrates with popular frameworks like TensorFlow and ResNet for ease of implementation.
[DEPRECATED] Self-Supervised Extrinsic Calibration using 3D Spatial Transformer Networks
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Eliminates the need for manually labeled calibration data by using a point cloud distance loss, reducing cost and time for large datasets.
Supports distributed training across multiple GPUs, as shown in the training commands, enabling faster model optimization.
Includes pretrained ResNet-18 weights and trained CalibNet parameters, allowing for quick start or fine-tuning without training from scratch.
Uses 3D spatial transformer networks to learn the calibration transformation in a single, trainable module, simplifying the alignment process.
The repository is marked as deprecated with no active support, directing users to an unofficial PyTorch implementation, which hinders reliability.
Requires specific versions like TensorFlow 1.3, CUDA 8.0, and CUDNN 7.0.1, with manual compilation of TensorFlow operators, making installation cumbersome.
Key components like evaluation code and iterative realignment models are not yet uploaded, limiting practical testing and deployment.