Take a video by bike and reconstruct 3D data from the city

[gkiavash]

• [x] Take test video
• [x] Create sample 3d points

In the next step:

• [x] Take a new test video with distorted config
• [x] undistort the images OR
• [x] tell COLMAP the distortion parameters from #7 and use fixed camera params
• [x] Apply pixel-perfect refinement

[gkiavash]

• Test video: 10 minutes, 1080p, 60 frames/sec
• Video to frames: every 10 frames, means 6 images per second
• Only the first 20 seconds are used
• Final dataset: 120 images, resolution 1920*1088, ~3MB size each image
• foggy weather, around 8AM

Output:

• Final database size: 225MB
• ~732k points
• ~2 hours to be processed

Sparse reconstruction:

Dense reconstruction:

[gkiavash]

	Demo video	Final udistorted video
Camera	1080p, 60f, linear	2.7k, 60f, wide
Video captured	~3GB, ~9 mins	~10.3GB, ~25 mins
Frames to COLMAP	Every 15 frames, 1920×1080, frame size: ~1MB	Every 15 frames, 2704×1538 , frame size: 1.4MB
Number of possible frames	2160	6000
Total size of possible frames	~2.1GB	~8.4GB
Number of used frames	497, 2 mintues	-
Total size of used frames	~500MB	-
Running COLMAP	num of features and matches limited to 4096 for RAM limit	-
Total size of sparse + db	~400MB	-
Number of sparse points	60k	-
Refine with pixel-perfect	impossible because of low gpu memory	-
Dense reconstruction(depth maps)	size: ~34.1GB, time: more than 5 hours	-
	Depth maps are created. But couldn't calc final fused.ply because of the lack of RAM, colmap crashed	-
Computer config	8GB ram + 10 swap(hdd), Nvidia 1650 4GB	-

[gkiavash]

Demo video:

[gkiavash]

• Test video: 20 seconds, 2.7k, 60 frames/sec, wide and distorted
• Video to frames: every 20 frames, means 3 images per second
• Sunny weather, around 8AM

Output:

• Final database size: 320MB
• Uncalibrated camera and distorted images
• Sparse Reconstruction, refined with pixel perfect: ~100k points
• Dense Reconstruction, NOT refined: ~400K points

• After adding calibration params and refinement with pixsfm, over 3 million points were detected in dense reconstruction. But, the point cloud is not meaningfull: