Citation Context

...uce the number of measurements taken of each point, keeping only the most useful measurements of that point. For a point p, bundle adjustment builds the information matrix V (we adopt the notation of =-=[5]-=-, Appendix 6.6): V = X j Vj = X j wjB T j Σ −1 j Bj where Bj is the point’s 2×3 projection Jacobian (w.r.t. point position) in the jth keyframe, with corresponding measurement noise Σ−1j and wj is the...

Citation Context

... from the iPhone are received at a resolution of 240×320 pixels, from which five pyramid levels (up to L4 at 15×20 pixels) are generated. We require that new features be valid Shi-Tomasi (ST) corners =-=[12]-=- in at least all of L0,L1,L2. This requirement allows efficient corner extraction, since exhaustive ST-corner extraction is only required in the (relatively small) L2 image. Each maximal corner is pro...

Citation Context

...one’s position in a previously unknown environment. In particular, we attempt to implement a keyframe-based SLAM system on the Apple iPhone 3G. Previous experience with the keyframe-based PTAM system =-=[6, 7]-=- on the PC has shown that it is able to generate room-sized maps and track a moving camera, while (crucially) offering a high level robustness to rapid motion, poor imaging conditions, and inevitable ...

Citation Context

...se of a large number of point measurements per frame. To generate such a number of points, every possible landmark is inserted into the map: For each of four image pyramid levels, every detected FAST =-=[11]-=- corner which can be triangulated is added. This can be wasteful, as a single physical point may have map duplicates at a number of scales. We use a different approach here, and require that features ...

Citation Context

...n of an initial stereo pair, along with point correspondences, from which an initial map is generated either using five-point-pose [13] or, in later versions of the system, a homography decomposition =-=[4]-=-. We have since found that requiring users to provide two initial keyframes is problematic, as some will not understand stereo baseline requirement and attempt to initialise the system using pure rota...

Citation Context

...splay. Even though the computational power available on phones still lags far behind desktop PCs, some researchers have already demonstrated that it is sufficient to perform real-time visual tracking =-=[8, 10, 14, 15]-=-. In particular, we are encouraged by [15]: Wagner et al. demonstrate that by carefully adapting, refining and combining established methods, phones are capable of extracting FAST corners, detecting S...

Citation Context

...er-based systems have also been demonstrated on phones. A recent example attempting to reduce the area taken up by the physical marker is [14], but demonstrations of mobile marker tracking go back to =-=[16]-=- in which ARToolkit markers were detected on a PDA and [8] who develop their own marker on an early camera phone. (Today FPUs, OpenGL acceleration and faster CPUs have greatly reduced the engineering ...

Citation Context

...ational cost of adding them to the map is prohibitive on the iPhone. On the other hand full-frame rotation estimation, which aligns successive video frames by direct 3-DOF efficient second-order (ESM =-=[2]-=-) alignment on the coarsest (15×20pixel) pyramid level, is very fast and we use this method here. This effectively acts as a gyroscope, which the iPhone physically lacks. PTAM on the PC tracks point f...

Citation Context

...om a planar scene PTAM on the PC must be bootstrapped by the provision of an initial stereo pair, along with point correspondences, from which an initial map is generated either using five-point-pose =-=[13]-=- or, in later versions of the system, a homography decomposition [4]. We have since found that requiring users to provide two initial keyframes is problematic, as some will not understand stereo basel...

Citation Context

... alone, it is clear that some compromises need to be made to perform bundle adjustment on the phone: A numerical reduction in map size is required. The scalability of bundle adjustment is detailed in =-=[3]-=-. Given a map of N points and M keyframes, with each point observed by T keyframes, different parts of bundle adjustment scale with different quantities: projections, differentials and back substituti...

Citation Context

...splay. Even though the computational power available on phones still lags far behind desktop PCs, some researchers have already demonstrated that it is sufficient to perform real-time visual tracking =-=[8, 10, 14, 15]-=-. In particular, we are encouraged by [15]: Wagner et al. demonstrate that by carefully adapting, refining and combining established methods, phones are capable of extracting FAST corners, detecting S...

Citation Context

...splay. Even though the computational power available on phones still lags far behind desktop PCs, some researchers have already demonstrated that it is sufficient to perform real-time visual tracking =-=[8, 10, 14, 15]-=-. In particular, we are encouraged by [15]: Wagner et al. demonstrate that by carefully adapting, refining and combining established methods, phones are capable of extracting FAST corners, detecting S...

Citation Context

...AM system (where, typically, the assumption of a global shutter is made). While some authors have demonstrated that a rolling shutter can be exploited to directly measure velocity from a single image =-=[1]-=-, this was done in the context of fast high-res cameras, small accelerations, and known models; here, we merely aim to reduce the negative impact of the rolling shutter on map accuracy. This is done b...

Citation Context

...splay. Even though the computational power available on phones still lags far behind desktop PCs, some researchers have already demonstrated that it is sufficient to perform real-time visual tracking =-=[8, 10, 14, 15]-=-. In particular, we are encouraged by [15]: Wagner et al. demonstrate that by carefully adapting, refining and combining established methods, phones are capable of extracting FAST corners, detecting S...

Citation Context

...ully measured is re-triangulated against the original keyframe and immediately inserted into the map with standard Cartesian coordinates. This means that in contrast to most MonoSLAM systems (such as =-=[9]-=-, which this approach resembles) the inverse-depth information filter is used purely for data association: Any information gathered from the intermediate frames is discarded once a second keyframe obs...

Citation Context

...tion described in §3.1: Instead of always sourcing template pixels from the same pyramid levels and varying the search level (as in [6]), here we can fix search level and vary the source level (as in =-=[17]-=-) to compensate changes in scale as the camera moves, and still be assured a trackable template. When the camera is rotating quickly even the L1 image becomes too blurred. When this is predicted by th...