There are some issues with dynamic object culling

[SSSSSSStar]

As mentioned in the paper, it uses semantic information to perform dynamic object removal during the map update phase, and provides a formula for the semantic consistency penalty function. However, I cannot find the relevant code for dynamic object removal in the code. Could you please tell me which files are included?

[jbehley]

A bit late, but maybe it helps someone else. Sorry :/

The dynamic object removal is only implicitly happening by putting a penalty on the association in the computation of the Jacobian:

semantic_suma/src/shader/Frame2Model_jacobians.geom

Lines 143 to 158 in 531954d

	// use semantic information during ICP
	float data_label = texture(semantic_data, texCoords).x * 255.0;
	float data_label_prob = texture(semantic_data, texCoords).w;
	float model_label = texture(semantic_model, idx).x * 255.0;
	if( model_label == car.w || model_label == bicycle.w ||
	model_label == bus.w || model_label == motorcycle.w||
	model_label == truck.w|| model_label == other_vehicle.w||
	model_label == person.w||
	model_label == bicyclist.w || model_label == motorcyclist.w)
	{
	if(round(data_label) != round(model_label))
	weight *= (1 - data_label_prob);
	else
	weight *= data_label_prob;
	}

And the actual integration of this into the map building (to remove conflicting labels) is happening here:

semantic_suma/src/shader/update_surfels.vert

Lines 187 to 195 in 531954d

	if(round(data_label) != round(model_label))
	{
	if( model_label == car.w || model_label == bicycle.w ||
	model_label == bus.w || model_label == motorcycle.w||
	model_label == truck.w|| model_label == other_vehicle.w||
	model_label == person.w||
	model_label == bicyclist.w || model_label == motorcyclist.w)
	penalty = 1.0;
	}

The penalty term then reduces the confidence and surfels below a certain confidence are then removed.