Report on Error Analysis & Synchronization (<--- click to view pdf):
---- Section I: Error Analysis of MarkIt
---- Section II: Error Analysis of Dual-camera recorder
---- Section III: Synchronization Among Sensors and Cameras
---- Section IV: Error Handling
---- Section V: Comparison with Other Fiducials
Video: A less animated video for our work (<--- click to view new video)
Rigid and articulated objects are common and important in everyday life. The research on visual analysis of rigid and articulated objects, such as pose estimation and tracking, is gaining increasing attention in both robotics and computer vision communities. However, building a large-scale real-world benchmark requires considerable manual effort.
To alleviate the burden of dataset construction, we propose a visual-inertial tracking system named VIRAT, which can accurately record the object pose and save much human labor. The VIRAT system consists of proposed wireless active markers MarkIt, a dual-camera recorder, and a data acquisition suite. To demonstrate the wide usage of the VIRAT system, we apply it to different application scenarios, including fixed and portable settings, and also working with robots.
Besides, we construct an object pose dataset BAO to demonstrate that VIRAT can democratize the recording pipeline. We perform experiments on BAO, showing that MarkIt barely affects the performance of perception models, making it capable of providing data to train models that are generalizable to markerless application scenarios.
The MarkIt has double-sided PCB with IMU andLED on one side and Wi-Fi transceiver on the other side, and the battery. The LED can be connected horizontally or vertically.
The MarkIt uses an ESP32S chip made by Espressif and a HI229 inertial measurement unit made by HIPNUC. It supports the connection of a 3.7v lithium battery and an LED through 2Pin connectors.
The MarkIt can be used along with mainstream fiducial markers (e.g. apriltags)
PCB Design: Download
IMU Specs: HI229 Specs (PDF version: HI229 Specs)
The VIRAT system has a dual-camera recorder which is composed of an Intel Realsense L515 and D435. In the portable setting, we manage to minify the data acquisition suite by adopting a mobile AP and router for network connection and a laptop to control the MarkIt and the dual-camera recorder. Thus, it can be wrapped in a backpack. The mobile power supply can support the whole system working for 2 hours.
Code for LED detection: Download
Software for Annotation: Coming soon
Dataset Sample Download: robotflow/bao
Baseline models: Coming soon