ORB-SLAM2 and COLMAP
This module focused on gaining familiarity with visual SLAM systems, specifically ORB-SLAM2 and COLMAP, with an emphasis on monocular SLAM. The primary objective was to collect and evaluate small datasets that demonstrate both the strengths and limitations of ORB-SLAM2, while using COLMAP reconstructions as a ground truth reference for comparison.
To investigate performance, various camera settings (frame rate, resolution, and exposure) were tested to determine how these parameters affect reconstruction quality and robustness.
Outdoor Track Sequence
The first experiment involved walking around a running track, a scene expected to be ideal for SLAM due to its simple path and open visibility. However, the low visual diversity — repetitive colours, distant trees, and few unique features — caused ORB-SLAM2 to fail completely, unable to maintain sufficient feature tracking.
In contrast, COLMAP successfully completed the reconstruction, highlighting its ability to exploit high-resolution imagery to identify and match subtle scene features.
Indoor Navigation Sequence
The second test took place in an indoor university setting, where the goal was to create a controlled environment with known feature points. Initially, both systems struggled due to low texture areas and large frame-to-frame movements. To improve results, we manually increased scene texture by placing magazines, objects, and pool balls within the environment.
We also mounted the camera on a wheeled chair for smoother motion, reducing motion blur and frame jumps. These adjustments led to excellent reconstructions for both ORB-SLAM2 and COLMAP, producing accurate, consistent maps with strong feature correspondence.
Outdoor Driving Sequence
Finally, a KITTI-style driving dataset was captured by mounting a stabilized camera to a car roof and driving through a suburban neighbourhood.
This test showcased ORB-SLAM2’s strongest performance, maintaining stable feature tracking despite objects moving both toward and across the camera’s field of view. The distinct geometry and texture of the environment (buildings, road markings, and vehicles) produced a clean, high-quality reconstruction.
💡 Key Takeaways
- Gained hands-on experience with ORB-SLAM2 and COLMAP for monocular visual SLAM.
- Conducted experiments comparing frame rate, resolution, and environmental texture on reconstruction quality.
- Discovered the strengths of COLMAP in static high-resolution contexts versus ORB-SLAM2’s real-time adaptability.
- Learned to capture and prepare datasets suitable for benchmarking SLAM performance.