Robust wide baseline pose estimation from video

Abstract

Abstract-Robust wide baseline pose estimation is an essential step in the deployment of smart camera networks. In this work, we highlight some current limitations of conventional strategies for relative pose estimation in difficult urban scenes. Then we propose a solution which relies on an adaptive search of corresponding interest points in synchronized video streams which allows us to converge robustly towards a high-quality solution. The experiments are performed using a manually annotated ground truth of a large scale scene exhibiting significant depth and perspective variation, uniform areas, repetitive patterns and homogeneous dynamic elements. The results show a fast and robust convergence of the solution, and a significant improvement, compared to single image based alternatives, of the RMSE of ground truth matches, and of the maximum absolute error.

Wide baseline pose estimation from video with a density-based uncertainty model

“…7a shows the spatial distribution of the symmetric geometric error on the left image. For each bucket of the image, we highlight the average error of the estimation with respect to the matches drawn from All-matches [44] Proposed method Fig. 8 RMSE and Max geometric error by applying the All-matches strategy, the method in [44] and our algorithm on 1-2 camera pair of Regents Park dataset.…”

“…For each bucket of the image, we highlight the average error of the estimation with respect to the matches drawn from All-matches [44] Proposed method Fig. 8 RMSE and Max geometric error by applying the All-matches strategy, the method in [44] and our algorithm on 1-2 camera pair of Regents Park dataset. Our selection is more reliable, and we are able to improve the initial estimation significantly and robustly, with a lower RMSE and less oscillations than [44] the ground-truth points at that location.…”

“…8 RMSE and Max geometric error by applying the All-matches strategy, the method in [44] and our algorithm on 1-2 camera pair of Regents Park dataset. Our selection is more reliable, and we are able to improve the initial estimation significantly and robustly, with a lower RMSE and less oscillations than [44] the ground-truth points at that location. While approaching the area lacking inliers, we note the presence of high errors, which makes the single image pair approach unadapted for fitting the entire image space.…”

“…In our preliminary work [44], we defined the notion of well-constrained regions by using a fundamental concept introduced in the field of data clustering with noisy data [13]. In [13], a point q is considered a core point if, given two parameters and Min Pts, | N (q) |≥ Min Pts, where N (q) is the set of points at a distance lower than from q.…”

“…4) -we introduce an algorithm for evaluating the local uncertainty as a nonlinear mapping of the observation density. Our density estimation is a continuous extension of the density-based spatial clustering which was used in our initial work [44] (Sect. 5) -we propose an algorithm variant which refines an existing pose, and which is more suitable for stereo systems requiring frequent extrinsic calibrations (Sect.…”