Author: Engel & Cremers
Year: 2016
- Do not depend on keypoints & descriptors => acheive real time
- Indirect method: pre processing of raw measurment to compute intermediate values
- Direct methods use the raw sensor measurements
- Direct => photometric error, indirect => geometric error
- Dense method exploit the connectedness of pixels => a geometric prior for smoothness
- Benefit of direct formulation = a pixel does not need to be recognizable by itself
- Dense formulation => correlation between pixels that makes the probability distr intractable
- Common features descriptors and detectors invariant to photometric variations
Photometric calibration:
- non linear response function and lens attenuation (vignetting) => impose photometric correction to each image
Model formulation:
- photometric error of a point is a weighted SSD (?) on a neighbourhood of pixels
- weighting based on brightness funciton parameters, time exposure and pixel gradient
- inverse depth parametrization of points
Optimization:
- Windowed optimization in a gauss newton scheme
- "First evaluate Jacobians" = evaluating geometric and photometric jacobians at x=0 as these are smooth functions
- Marginalization of old variables with the Schur Complement
- Marginalization = "elimination of states in nonlinear optimization"
VO front-end:
- determining the set of points w/ outlier rejection and occlusion detection
- good initialization is necessary so that the linearization remain good (rule of thumb: linearization valid in 1-2 pixel radius)
Frame management:
- 7 active KF all the time
- Initial frame tracking is performed with two frame direct image alignement with constant velocity model
- lots of KF in the beginning then it is sparsified
- KF creation with optical flow and exposure time values
- marginalization strategy: if less than 5% of keypoints are shared with the last KF and wrt to the euclidean distance
- Np = 2000 active points in all active KF, at each KF: -> 2000 points are selected (well distributed, with high gradient)
-> these are tracked in subsequent frames (on epipolar line, minimizing photometric error) to compute a coarse depth value
-> then 2000 points are selected on all active KF and added to optimization pb
Analysis:
- Direct methods models photometric noise ie additive noise on pixel intensity while Indirect methods models geometric noise ie noise on (u,v) coords of pixels
- Direct method works well with camera designed for computer vision as the geometric noise is corrected
- Hihgly non convex pb du to the image parameters included in the cost function