Object recognition from local scale-invariant features
Citations
202 citations
201 citations
Cites methods from "Object recognition from local scale..."
...Each image region is represented as a 72 dimensional SIFT [17] descriptor....
[...]
201 citations
Cites background from "Object recognition from local scale..."
...At the end of this merging process, an object database with fused features is formed, where the descriptors are taken from each of the N − 1 plain images, and their locations are transformed to the reference image coordinates (see Fig....
[...]
...A comparative knowledge-acquisition system appears in [16], consisting of several object recognition modules that represent a car image viewed from the rear, such as a window, tail lights, and so on, based on color recognition....
[...]
201 citations
Cites background from "Object recognition from local scale..."
...Scale Invariant Feature Transform (SIFT) descriptors [18] are computed on that region....
[...]
201 citations
Cites methods from "Object recognition from local scale..."
...Recent methods for matching features based on image feature invariants, such as Lowe (1999), are good candidates for this task....
[...]
...For this problem, correlation tracking techniques, such as that of Lucas and Kanade (1981), have often been used in the SFM literature, and we have used Lucas–Kanade with manual correction in the experiments we describe in Sections 7, 8, and 9....
[...]
...As described in Section 11.3, this is future work and we expect to address this issue by exploiting recent work on image feature invariants (e.g. by Lowe 1999) and data association approaches....
[...]
...This is particularly true as the time propagation variances, which specify how close we expect camera positions at adjacent times to be, are allowed to grow, easing the assumption of motion smoothness, which the batch method does not incorporate. Our batch and recursive image-and-inertial algorithms are not as closely related in this way. For instance, the batch algorithm estimates the sensor position only at the time of each image, while the recursive algorithm estimates the sensor position at the time of each image and each inertial measurement. More importantly, the multirate design of the recursive algorithm implicitly requires the assumption of smoothness in the angular velocity and linear acceleration across measurement times to exploit inertial and image measurements acquired at different times. Consider, for example, an inertial measurement followed by an image measurement. If the angular velocity and linear acceleration time propagation variances are high, as required in scenarios with erratic motion, then the state prior covariance that results from the time propagation between the inertial and image measurement times grows quickly and only loosely constrains the image measurement step estimate. So, the filter is free to choose an estimate at the time of the image measurement that is quite different than the state estimate that resulted from the inertial measurement step. On the other hand, motion smoothness is not a critical assumption for combining measurements taken at different image times, as it is in the recursive image-only filter, because these estimates are related through the threedimensional positions of the points observed at both times. Our batch algorithm, including the integration functions Iρ , Iv, and It that integrate inertial measurement between image times, neither requires nor exploits such an assumption, so the batch algorithm is stronger than the recursive algorithm in the presence of erratic motion. In our experience, robustness to erratic motion is more valuable in practice than the assumption of motion smoothness, and Tomasi and Kanade (1992) drew a similar conclusion in the context of SFM....
[...]
References
[...]
5,672 citations
4,310 citations
2,037 citations
1,756 citations
"Object recognition from local scale..." refers background or methods in this paper
...This allows for the use of more distinctive image descriptors than the rotation-invariant ones used by Schmid and Mohr, and the descriptor is further modified to improve its stability to changes in affine projection and illumination....
[...]
...For the object recognition problem, Schmid & Mohr [19] also used the Harris corner detector to identify interest points, and then created a local image descriptor at each interest point from an orientation-invariant vector of derivative-of-Gaussian image measurements....
[...]
..., Schmid & Mohr [19]) has shown that efficient recognition can often be achieved by using local image descriptors sampled at a large number of repeatable locations....
[...]
...However, recent research on the use of dense local features (e.g., Schmid & Mohr [19]) has shown that efficient recognition can often be achieved by using local image descriptors sampled at a large number of repeatable locations....
[...]
1,574 citations
"Object recognition from local scale..." refers methods in this paper
...[23] used the Harris corner detector to identify feature locations for epipolar alignment of images taken from differing viewpoints....
[...]