Author: Kasper
Year: 2019
- 160 minutes of video/IMU with 1500/4500/9000 lumens
- start and end at the same position -> accumulated drift as a metric
- utility of LiDAR reduced in long hallways or tunnels
- light source violates brightness consistensy assumption made in most direct methods
- use a light model
- vicalib for calibration of the sensor rig
- photometric calibration is provided
- uses SplineFusion to estimate temporal offset between each sensor and gt measurements
- captures of targets to produce light calibration
- basic calibration of light is given by $$ I = \frac{\alpha L \cos \theta}{d^2} $$
- details of metric used for loop closure
- OKVIS is the best, ORB SLAM failed under bad illumination conditions
Commentaires:
- excellent papier dataset