Supplementary Material for Exploiting Object Similarity in 3 D Reconstruction

Chen Zhou,Fatma Güney,Yizhou Wang,Andreas Geiger

2015-01-01

Abstract:In this supplementary material we first provide details of our experimental setup and the inference process with respect to the individual optimization blocks. Furthermore, we provide additional quantitative and qualitative results: First, we show the full plots corresponding to Figure 5 in the main paper. Second, we visualize the accuracy and completeness of the individual methods. And finally, we qualitatively compare the results of the proposed approach to the initial reconstruction baseline on novel sequences. 1. Experimental Setup For capturing our dataset and the 3D ground truth, we used a Velodyne HDL-64 laser scanner, running at 10 fps. For synchronization, we attached a hardware trigger to the scanner (reed contact) which triggers all cameras at the moment the scanner is facing forward. Each scan returns the timestamp of each point which is used to unwarp the point cloud according to the vehicle motion obtained using a high-precision OXTS 3003 RTK GPS/IMU unit (2cm accuracy). The cameras have a global shutter and are in perfect sync. The average driving velocity was 40 km/h. 2. Inference This section provides details on optimizing the individual blocks in our energy function argmin π,φ,X,V,k N ∑ i=1 Ψi(π,φ,X,V,k) (1) where Ψi(·) = ψshp(πi, φki ,xi) + λscale ψscale(si,vki) +λreg ψ (i) reg(πi) (2) The reader is referred to section 3.3 of the paper for the definition of each term. 2.1. Block {π,V} We optimize Eq. 1 with respect to {π,V} via gradient descent. For clarity of exposition we consider a single term Ψi, dropping the observation index i. First, we note that the derivatives of ψscale(s,vk) and ψreg(π) with respect to {π,V} are readily given due to their simple quadratic form. The derivative of ψshp(π, φ,x) with respect to π is given by ∂ψshp(π, φ,x) ∂π = ∑ p∈Ω (φ(p,x)− d(π(p))) ∂w ∂π − 2w(π(p))(φ(p,x)− d(π(p)) ∂d ∂π (3)