A Video Dataset for Feature-Poor and Low-Light AR Tracking Benchmarking: ORB vs AKAZE
Description
This dataset contains 36 video sequences (.mp4 format) captured at 10, 100, and 500 Lux. It is designed to evaluate the visual drift (RMSE) and tracking breakdown points of ORB and AKAZE algorithms within the Unity AR Foundation framework under various kinematic movements (static, panning, rotation) and surface textures.
Files
Steps to reproduce
1. ENVIRONMENT & PHYSICAL SETUP: Target Placement: Place the retail packaging box (Lenovo M210 Gaming Mouse) precisely at the center of the designated floor surface. This serves as the initial ground truth anchor for the tracking system. Photometric Control: Utilize an adjustable overhead lighting system. Employ a calibrated digital Lux Meter to measure and verify the light intensity at the object's surface center, ensuring constant values of 10, 100, or 500 Lux respectively. 2. DEVICE HARDWARE CONFIGURATION: Capture Settings: Utilize an Android-based mobile device camera. Configure the system to capture video streams at a resolution of 1920x1080 (Full HD) with a fixed, constant frame rate of 30 FPS. Raw Data Preservation: Disable all Electronic Image Stabilization (EIS), auto-gain control, and post-processing software filters. This is critical to maintain raw sensor noise profiles and prevent distortion of the original spatial features. 3. KINEMATIC DATA ACQUISITION: Static Sequences: Secure the device onto a rigid mechanical tripod. Record the target for a 60-second duration to capture baseline sensor noise and evaluate potential static visual drift. Panning (Translation) Sequences: Execute smooth, linear forward-and-backward camera motions along a controlled track. This process is essential to evaluate the scale-invariance capabilities of the descriptors under varying target sizes. Rotational Sequences: Perform manual multi-axis angular rotations around the camera's principal axis. These sequences are designed to introduce perspective shifts and natural motion blur, challenging the orientation estimation accuracy of the algorithms. 4. REPLICATION & CONTROL TRIALS: Statistical Sampling: Perform a minimum of three (3) independent replication trials for every matrix combination (Lux x Motion) to allow for robust two-way ANOVA statistical testing. Topology Baseline: For comparative background analysis, include separate recordings over a featureless tile surface (HOMOGEN.mp4) and a high-frequency repeating fabric pattern (TEKSTUR files). This aids in identifying the tracking system's lower-bound failure threshold.