Optimize network structure and loss function. Compare and evaluate output depth map.

• Observing with eyes, I think my depth map is better than monodepth2.
• But in the evaluation, my result is worse than monodepth2.
• For the same network, the evaluation value of the model trained by myself is worse than the … Continue Reading ››
  

1. Learn to write article with LATEX.
2. Optimize the network by joining different edges

Continue Reading ››

I trained a network to predict edge, but the result is not good. Professor Zhang suggest me to use edge calculated from traditional method. I think this is a good suggestion.

In the past, I just pay attention on the method to achieve a SLAM system. Today, Professor Zhang suggest … Continue Reading ››

1. In order of reduce the amount of parameters, changing some parts of depthnet in Monodepth2, and train the network with the data set from KITTI. But the result is not so well as Monodepth2.

Continue Reading ››

1. reproduced and trained DepthNet (a combination of CONV and LSTM). But what surprised me was that the memory usage was too large. So I plan to give up using LSTM.
2. MobileNet V1: Using depthwise convolutions and 1X1 convolutions instead of standard convolution to reduce the numble of parameters. And introducing two hyper-parameters (width … Continue Reading ››
   

very sorry a emergency happened, next week I will do more work.

1. monodepth2: Digging Into Self-Supervised Monocular Depth Estimation.

• Using the minimal appearance loss instand of average appearance loss, which can address the problem of occlusion pixel in monocular depth prediction.
• Proposed a auto_mask method to ignore pixel without camera motion.
• calculating the loss of each scale, and downample the single image … Continue Reading ››
  

1. train CNN_SLAM's network on NYU Depth Dataset V2, and test on NYU Depth Dataset V2 and TUM RGB-D SLAM Dataset. but it performed well on NYU , and not so well on TUM.

2. DVSO: Leveraging Deep Depth Prediction for Monocular Direct Sparse Odometry. this paper proposed … Continue Reading ››
   

read the paper(D3VO) published by the group of the TUM in 2020.

• the author proposed a self-supervised monocular depth estimation network;
• by predicting the brightness transformation parameters to align the brightness of source and target image;
• by using virtual stereo term to incorporate the predicted pose into graphic optimization.