Permalink
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
27 lines (22 sloc) 2.27 KB

Geometry-Aware Learning of Maps for Camera Localization

S Brahmbhatt, J Gu, K Kim, J Hays, J Kautz, CVPR, 2018

Summary

  1. MapNet Proposed to represent maps as a DNN : Learns the map representation directly from input data; flexibility to fuse multiple sensory inputs and to improve over time using unlabeled data.
  2. Contributions:
    • MapNet shows how the geometric constraints between pairs of observations can be included as an additional loss term in training. Constrain sources: pose constrian from VO, translation constrain from GPS, rotation constrain from IMU.
    • PoseNet et al are offline methods, the learned DNNs are fixed after training. MapNet+ can continuously update the DNN weights (i.e., maps)
    • Proposed a new parameterization for camera rotation which is better suited for deep-learning based camera pose regression.
  3. Proposed Approach:
    • Camera Pose Regression with DNNs
      • use ResNet-34 and modify PoseNet by introducing a global average pooling layer after the last conv layer, followed by a fc layer with 2048 neurons, a ReLU and dropout with p = 0:5. This is followed by a final fc layer that outputs a 6-DoF camera pose.
      • propose to parameterize camera orientation as the logarithm of a unit quaternion.
    • MapNet: Geometry-Aware Learning Data flow for our proposed algorithms.
      • MapNet enforces geometric constraints between relative poses and absolute poses in network training. MapNet+ fuses other inputs such as visual odometry to update maps with self-supervised learning. MapNet+PGO performs PGO at testing time to further improve accuracy.
    • MapNet+: Update with Unlabeled Data
      • MapNet+ fuses additional data (IMU GPS) to update the weights of MapNet with self-supervised learning : fine-tune a pre-trained MapNet by minimizing a loss function that consists of the original loss from the labelled dataset D and the loss from the unlabeled data T .
    • MapNet+PGO: Optimizing During Inference MapNet+PGO fuses the absolute pose predictions from MapNet+ and the relative poses from VO using pose graph optimization (PGO) to get smooth and globally consistent pose predictions.

Strengths / Novelties

MapNet+ use unlabeled videos and multiple sensory input.