Why initializer for monocular camera only matches features at octave 0?

[mitul93]

Hi

I was tracing the code for monocular map initialization and found out that area::match_in_consistent_area() only compares the features of current frame and previous frame at octave 0.

https://github.com/stella-cv/stella_vslam/blob/main/src/stella_vslam/match/area.cc#L22

I tried to comment the check and the system initialized faster with much more initial key points.

Is there a reason to use only features at octave 0 during initialization?

[ymd-stella]

Initializer seems to be implemented under the assumption that octave 0 is used.
This can be found in the following line:

stella_vslam/src/stella_vslam/initialize/perspective.cc

Line 52 in f94dbda

const float sigma = 1.0f;

sigma means standard deviation of measurement error.
When comparing features of different scales, appropriate sigma should be used from the table calculated below.

stella_vslam/src/stella_vslam/feature/orb_params.cc

Lines 53 to 61 in f94dbda

	std::vector<float> orb_params::calc_level_sigma_sq(const unsigned int num_scale_levels, const float scale_factor) {
	float scale_factor_at_level = 1.0;
	std::vector<float> level_sigma_sq(num_scale_levels, 1.0);
	for (unsigned int level = 1; level < num_scale_levels; ++level) {
	scale_factor_at_level = scale_factor * scale_factor_at_level;
	level_sigma_sq.at(level) = scale_factor_at_level * scale_factor_at_level;
	}
	return level_sigma_sq;
	}

initialize/bearing_vector.cc does not use sigma, so if equirectangular images are used as input, the check may be removed.

[mitul93]

Thanks for explanation @ymd-stella