3D_Drone_Reconstruction

AIM: 2D to 3D reconstruction of telecommunication tower.

VisualSFM

A completely integration of 3D reconstruction software:

• VisualSFM
  • Easy to run:
    • download feasible version from the website
    • download dependency: CMVS/PMVS by Pierre Moulon
    • simply follow the guide in the website

SFM

We implemented SFM using MATLAB. For feature detection SURF was used.

Feature Detectors

Get feature points, we used Lowe's program but failed to generate feasible solution with complex background and blur. Trying alternatives now:

• SIFT

  • Lowe's binary
  • OpenSIFT
    • Download the repo with git submodule
    • Compile

• SURF

• Minimum Eigenvalue Algorithm

• BRISK

Unet

We implemented a Unet model to segment out tower from images with complex background.

Image Augmentation

130 masks were manually created and augmented. Images were rotated flipped and blurred to increase dataset size.

Unet model

Total images: 4000

Total masks (labels): 4000

Training loss: 0.2660

Training accuracy: 88.44%

Validation loss: 0.2744

Validation accuracy: 87.84%

Test accuracy: 87.93%