Author: Joseph Ortiz & Andrew Davison
Year: 2020
- IPU = new calcul unit for AI with massive parallelisation
- GPB = general case of loopy belief prop.
- Not competitive on CPU but take
- BP = algorithm for computing marginal from joint probability
- When distributions are Gaussians, LGBP converges to the correct marginal
BA factor graph:
- 2 factors: measurement factor and prior factor
- prior factor = comparison to a prior distribution
- prior errors are required to set scale
- linearization of the projection function
- parametrization with information matrix and vector for measurement model
GPB:
- variable nodes update their belief by taking a product of incoming messages from adjacent factors
- synchronous scheduling: all factor nodes relinearise and send messages at each iterations
- message damping to stabilise the convergence
Results:
- Metrics to compare BA is ARE (average reprojection error)
Pas appris grd chose de nouveau, revoir la notion de marginalisation. Idée appliquer GPB à la fusion de capteurs