Summary of the results obtained for modeled and raw images.
Analyze the impact of modidy the number of samples in training.
Train a new LSTM net with different motion types and DOF to improve results.
Train a ConvLSTM with all motion types and DOF on raw samples.
Extend the pixel area to try to improve the CNN+LSTM results.
Train a combination of CNN with LSTM net with different motion types and DOF.
Train a CNN with all motion types and DOF on raw samples.
Analyze the impact of increasing the time gap.
Train a new LSTM net with all motion types and DOF to improve results.
Train a LSTM with all motion types and DOF on modeled images.
Train a MLP with all motion types and DOF on modeled images.
Generate new frame data using parabolic and sinusoidal dynamics .
Train different types of networks with modeled images.
Adapt the code to train and evaluate networks with modeled frames.
It is evaluated how a large increase in the number of samples affects the efficiency of the networks.
We focus on adapting the code to be able to read the dataset by batches and solve the memory error obtained with the large dataset.
We focus on training using raw images and linear movement with different types of networks and more samples.
New learning is tackled by valuing modeled frames and raw frames.
Compiling the work done so far and the results obtained to resume the project and establish a new starting point.
Linear motion is further complicated by letting the point start at a random height.
A new degree of freedom is added to the linear motion.
First contact with the created data.
The different types of data used for research in this thesis.