- It solves the Kidnapped-Robot problem by regressing poses of images that are part of a scene(used KITTI dataset's left image and poses ground truth).
- Paper Link
- It reuses VGG16 model coupled with Dense layers(4096, 4096) with a linear activation function(Wx+B) on the output, to regress poses of the input image scene.
- It has a custom loss function for rotation and translation componenets.
- L2_norm(y_pred, y_actual) for loss is used for both translation and rotation, with a Beta penalty factory for rotation only.
- Implement GTSAM and use this CNN based pose-regressor as a sensor along with other sensors such as GPS, IMU, etc for reliable odometry source.
- Add a feature where the model can regress both euler and quaternions depending on the input and out desired. Currently only outputs, x y z roll pitch yaw.
- Implement transfer learning so that we dont have to retrain the complete model and can fine tune it which is efficient.