End-to-End 6DoF Pose Estimation From Monocular RGB Images

We present a conceptually simple framework for 6DoF object pose estimation, especially autonomous driving scenarios. Our approach can efficiently detect the traffic participants from monocular RGB image while simultaneously regressing their 3D translation and rotation vectors. The proposed method 6D-VNet, extends Mask R-CNN by adding customised heads predicting vehicle's finer class, translation. It is trained end-to-end compared to previous methods. Furthermore, we show that inclusion of translational regression in joint losses crucial estimation task, where distance along longitudinal axis varies significantly, e.g., Additionally, incorporate mutual information between via modified non-local block capture spatial dependencies among detected objects. As opposed original implementation, weighting modification takes neighbouring into consideration whilst counteracting effect extreme gradient values. evaluate our on challenging real-world Pascal3D+ dataset 6D-VNet reaches 1st place ApolloScape challenge Car Instance task (Apolloscape, 2018), (Huang et al., 2018).