helga-helga / stereo-vision Goto Github PK

After diffusion is done, take the lightest vertex in each pixel: for each pixel (i, j) find disparity value d that gives the smallest value vertex_penalty_with_potentials(i, j, d). Minimum vertex weight for pixel is calculated in min_penalty_vertex() function. This function can be used but should be updated to return not only a vertex weight min_penalty_vertex, but also a disparity value d.

Then draw a matrix which size match a size of an input image and has disparities as values:

• dark pixel will denote that a pixel in an input image is far away (is deep)
• light pixel will denote that a pixel in an input image is close to the camera

• Investigate how to generate documentation
  https://doc.rust-lang.org/1.30.0/book/first-edition/documentation.html
  https://doc.rust-lang.org/rust-by-example/meta/doc.html
• Update comments for pgm_handler
• Update comments for penalty_graph
• Update comments for crossing_out_graph
• Update comments for epsilon_search
• Update comments for diffusion
• Update README.md specifying how to generate documentation and write comments

A problem can be either in diffusion algorithm or in energy calculating. But zero-problem penalty is constant.

After a batch of diffusion iterations, do crossing out with provided epsilon. When a task becomes consistent (crossing out returns non-empty graph), stop diffusion iterations.

• Investigate approaches to test functions that read or write into files (for example, https://crates.io/crates/tempdir)
• Write a test for pgm_reader()
• Write a test for pgm_writer()

Allow edges only where
d(left) <= d(right) + 1 for all horizontal neighbors (left, right), where d(left) is disparity of left pixel, d(right) is disparity of right pixel

Try to use 1 iteration of the ICM algorithm after diffusion algorithm instead of crossing out, as another primal heuristic.

• Investigate Travis CI for Rust
• Configure GitHub and Travis CI integration
• Add config file for CI
• Add status image in README.md

Zero disparity means that a pixel is infinitely far. It is the same as unknown distance.

TODO: create a separate penalty for zero disparity in lookup_table

Cover the next functions:

• neighbor_exists()
• neighbor_index()
• number_of_neighbors()
• approx_equal()

Big disparity means that two neighbor pixels are related to different objects.

For example, there is an object in the front of the wall. On the border between the object and the wall there will be a big disparity (consequently, a big penalty). And the far the wall is, the bigger is the penalty. It's normal situation, when objects has different distance from camera. But there is a big penalty for such a situation in a program.

TODO: If |d_1 - d_2| > c, then fix the penalty, so that it won't be dependent on d_1 and d_2,
where d_1 and d_2 are disparities of neighbor pixels and c is some constant.

Will the problem be still sub-modular?

Remove code duplicates: zero_penalty() and zero_one_penalty() compute penalty of scecific disparity maps. penalty() does the same. Just provide a disparity map to penalty() and use it

Unused -- where vertex doesn't exist (if j >= d doesn't take place)

• Rename penalty_graph module and structure to diffusion_graph
• Rename diffusion module to utils module
• Put Rust project into separate folder inside repository

Save graph and iteration number to file and then load it

An error could occur after changes in edge_penalty_with_potential() function (not all neighbors are used when calculating a penalty). Probably, the test should be updated.

---- penalty_graph::penalty_graph::test_penalty_four_pixels stdout ----
thread 'penalty_graph::penalty_graph::test_penalty_four_pixels' panicked at 'assertion failed: `(left == right)`
  left: `4.0`,
 right: `2.0`', src/penalty_graph.rs:393:9

Add tests for the next functions:

• create_array_of_epsilons()
• epsilon_search()

There is no test that checks if potentials are calculated correctly

• Integrate https://coveralls.io/ or https://codecov.io/ with GitHub
• Add execution of coverage report to .travis.yml (https://crates.io/crates/cargo-tarpaulin)
• Add status image to REAMDE.md

Add tests for the next functions:

• initialize_vertices()
• initialize_edges()
• initialize_with_epsilon()
• crossing_out()
• is_not_empty()
• vertices_exist()
• edges_exist()
• min_vertex_between_existing()
• cross_vertex()
• find_best_labeling()