Stereo-Matching

General pipeline: Cost Computation -> Cost Aggregation -> Disparity Optimization -> Disparity Refinement

-

1. Cost Computation:

Census cost = Local binary pattern -> Hamming distance

2. Cost Aggregation:

Refine the cost according to nearby costs. [Tips] Joint bilateral filter (for the cost of each disparty)

3. Disparity Optimization:

Determine disparity based on estimated cost. [Tips] Winner-take-all

4. Disparity Refinement:

Do whatever to enhance the disparity map.[Tips] Left-right consistency check -> Hole filling -> Weighted median filtering

-

Resuls

Tsukuba: bad pixel rate = 4.41%

Venus: bad pixel rate = 1.52%

Teddy: bad pixel rate = 9.39%

Cones: bad pixel rate = 7.85%