Modify keyframe's selection part in CNN_SLAM and upload to https://mountain.elcs.kyutech.ac.jp/zhou/cnn_slam.git.

About overlap ratio, calculating by judging whether points in keyframe are mapped in current frame.

1. Learning some relevant mathematical theories, including Quaternions, Lie-Algebras, Lie-Group;

2. LSD-SLAM:

• Track by calculating ξ∈se(3);
• optimization by calculating ξ∈sim(3). Solve accumulated scale-drift caused by long trajectories;
• Mean inverse depth is used to calculate distance of current frame and current keyframe so that every … Continue Reading ››
  

• After reading some paper, I find that most paper are depth-supervised or a combination of depth, pose and stereo. And I don't find any paper only using pose-supervised.
• The reason why not only using pose-supervised is that only considering large intensity gradient area will produce better results. And area with less texture will cause … Continue Reading ››
  

• reproduce the code of paper(DepthNet), and upload to https://mountain.elcs.kyutech.ac.jp/zhou/convlstm.git
• problems: there is no stride in FC_LSTM and LSTM, but it exist in CNN. I used interpolation to process C of pre_layer, but it may not be the best method. Maybe this is a point that can be improved in the future.

paper reading

1. DepthNet: A Recurrent Neural Network Architecture for Monocular Depth Prediction

• this paper is an extension of depth prediction scheme proposed by Eigen [6];
• Replace traditional CNN with ConvLSTM proposed by [31], so that the network can learn spatio-temporal information

関連課题題究:

Reproduce the code of each part in the paper(CNN_SLAM) based on pytorch, and upload to https://mountain.elcs.kyutech.ac.jp/zhou/cnn_slam/-/tree/master/zs_cnn_slam .The workflow of each part is show as bellow:

However, there are 3 problems that I … Continue Reading ››

先週の進捗

• Contribution of the paper: refined the CNN-predicted depth map associated to each key-frame via small-baseline stereo matching.
• Camera pose estimation: inspired by key-frame … Continue Reading ››