3D Reconstruction from accidental motion

Reconstructs a 3D scene by generating a dense depth map from a set of images capturing a scene.

The algorithm first uses KLT Tracking to track features across all images. We use Bundle Adjustment using the Ceres Solver and estimate the extrinsic camera parameters. Finally, we solve a smooth dense map using a DenseCRF formulation. Implemented as a part of the Computer Vision course at IIIT H.

Building Ceres Solver

• mkdir ceres-bin
• cd ceres-bin
• cmake ../ceres-solver-1.11.0/
• make -j4 or
• bash recompile.sh

Install dependencies and requirements

• pip install -r requirements.txt

Running KLT tracking and Bundle Adjustment

• cd src
• python main.py

Running DenseCRF

• python dense_crf.py --min_d 2 --max_d 4 --scale 2 --patch_rad 1 --iters 100 --p_std 3,3 --c_std 20,20,20 --wt 1 --max_p 0.15 --folder stone4 --nsamples 64 --show_wta

• If plane sweeping is already computed for the scene, add the flag --pc <path-to-npz>

Outputs

• cd outputs

Building DenseCRF package

• cd src
• chmod +x src/pydensecrf_setup.sh
• ./pydensecrf_setup.sh

Note: Make sure Cython is installed before running the above commands.

Dataset Location

• Download the dataset from here and store it in ./datasets

TODO

• KLT
• Filtering Good Points
• Bundle Adjustment
• CRF Energy Minimization
• Integration