Homography Decomposition Networks for Planar Object Tracking

Abstract

Planar object tracking plays an important role in AI applications, such as robotics, visual servoing, and visual SLAM. Although the previous planar trackers work well in most scenarios, it is still a challenging task due to the rapid motion and large transformation between two consecutive frames. The essential reason behind this problem is that the condition number of such a non-linear system changes unstably when the searching range of the homography parameter space becomes larger. To this end, we propose a novel Homography Decomposition Networks~(HDN) approach that drastically reduces and stabilizes the condition number by decomposing the homography transformation into two groups. Specifically, a similarity transformation estimator is designed to predict the first group robustly by a deep convolution equivariant network. By taking advantage of the scale and rotation estimation with high confidence, a residual transformation is estimated by a simple regression model. Furthermore, the proposed end-to-end network is trained in a semi-supervised fashion. Extensive experiments show that our proposed approach outperforms the state-of-the-art planar tracking methods at a large margin on the challenging POT, UCSB and POIC datasets. Codes and models are available at https://github.com/zhanxinrui/HDN.

Warped Convolution Networks for Homography Estimation

“…For the scale and rotation groups, γ represents the uniform scale, and θ denotes the in-plane rotation. As described in [14,38], the warping function w 1 for two parameters scale γ and rotation θ is defined as:…”

“…Henriques & Vedaldi [14] employ the warp function on convolution and implement two-parameter group equivariance, since there is utmost two dimensions independent possibility for an image. Zhan et al [38] recently decompose the homography into two groups and estimate them in order, which lost the equivariance for the residual parameters. Besides, it requires an additional homography estimator to find the corners.…”

“…We select the videos with perspective changes as our testing datasets. The state-of-the-art visual object trackers and planar object tracking approaches [38,40] show that it is easy to predict the offset of an object through the cross-correlation. Thus, we make use of it for parameter estimation in our framework.…”

“…For planar object tracking, we employ HDN [38] as our baseline method, which has three warp functions to predict b in addition to [t 1 , t 2 , γ , θ]. Besides, the perspective changes are small in the feature map.…”

“…All transformation penalties make use of the robust loss function (i.e., smooth L1) defined in [10]. As for POT, we employ the same classification and offset loss as HDN [38] for newly added parameters in x.…”