ORB-SLAM (纯Python实现) - TUM fr1_xyz 数据集测试
ORB-SLAM Baseline Experiment
This experiment tests a traditional ORB-based SLAM approach using pure Python + OpenCV as a baseline comparison to the deep learning-based CNN-SLAM approach with DINOv2 + LoFTR.
Implementation Overview
| Component | Method |
|---|---|
| Feature Extraction | OpenCV ORB (Oriented FAST and Rotated BRIEF) |
| Feature Matching | Brute-force matching + Lowe's ratio test |
| Pose Estimation | Essential matrix + RANSAC |
| Reconstruction | Keyframe-based triangulation |
| Interface | Follows the same BaseSLAM interface as CNN-SLAM for fair comparison |
Dataset Information
TUM RGB-D Dataset: fr1/xyz
| Property | Value |
|---|---|
| Original Frames | 798 |
| FPS | 30 |
| Processing FPS | 15 |
| Processed Frames | 399 |
| Resolution | 640 × 480 |
| Calibration | Official TUM intrinsics used |
Results Summary
| Metric | Value |
|---|---|
| Tracking Success Rate | 77.7% (311/399 frames) |
| Number of Keyframes | 264 |
| Reconstructed Map Points | 97,259 |
| Total Trajectory Distance | 35.87 meters |
Visualizations
1. ORB Feature Extraction Result (First Frame)
OpenCV's ORB detector extracts 1000 features from the first frame. Each green circle shows a detected keypoint with size proportional to the FAST response strength.
2. Reconstructed Trajectory and Point Cloud (3D View)
- Blue line = Estimated camera trajectory
- Green dot = Start position
- Red dot = End position
- Gray points = Reconstructed 3D map points (subsampled for visualization)
3. Top-Down View (X-Z Plane)
This view shows the bounding box of the trajectory matches the expected fr1/xyz dataset motion pattern.
ORB Feature Extraction
OpenCV's ORB detector extracts 1000 features from the first frame. Each green circle shows a detected keypoint with size proportional to the FAST response strength.
OpenCV's ORB detector extracts 1000 features from the first frame. Each green circle shows a detected keypoint with size proportional to the FAST response strength.
3D Trajectory & Point Cloud
Reconstructed camera trajectory (blue line) and sparse 3D point cloud (gray points) from ORB-SLAM.
Reconstructed camera trajectory (blue line) and sparse 3D point cloud (gray points) from ORB-SLAM.
Top-Down Trajectory View
XZ plane projection of the reconstructed trajectory shows the circular path shape.
XZ plane projection of the reconstructed trajectory shows the circular path shape.
Raw JSON Data
{"dataset": "TUM RGB-D fr1_xyz", "method": "ORB-SLAM (pure Python OpenCV implementation)", "input_frames": 798, "processed_frames": 399, "target_fps": 15, "success_rate": 77.7, "num_keyframes": 264, "num_map_points": 97259, "tracking_failures": 89, "total_distance": 35.869, "hyperparameters": {"n_features": 1000, "min_inliers": 20, "keyframe_thresh": 0.3, "match_ratio_thresh": 0.7}, "trajectory_bounds": {"X_range": 1.992, "Y_range": 1.99, "Z_range": 1.999}, "visualizations": [{"url": "/experiments/fr1_xyz_orb_slam/orb_features_first_frame.png", "title": "ORB Feature Extraction", "description": "OpenCV's ORB detector extracts 1000 features from the first frame. Each green circle shows a detected keypoint with size proportional to the FAST response strength."}, {"url": "/experiments/fr1_xyz_orb_slam/trajectory_pointcloud_3d.png", "title": "3D Trajectory & Point Cloud", "description": "Reconstructed camera trajectory (blue line) and sparse 3D point cloud (gray points) from ORB-SLAM."}, {"url": "/experiments/fr1_xyz_orb_slam/trajectory_pointcloud_topdown.png", "title": "Top-Down Trajectory View", "description": "XZ plane projection of the reconstructed trajectory shows the circular path shape."}]}
0 Comments
Sign in to add a comment