• About the result of past experiments, I make a summary. I think it may be because the data set is too simple, causing the network to be over-fitted before it is well-fitted. The reasons why I think so are :
  1. only the positive face is used for train;
  2. The result that training with … Continue Reading ››
     

• Do some experiments to test the influence of some parameters, including the number of illumination type, the number of image in each illumination, and the difference between illuminations.
• About the distribution of each image in combination, making a combination about the mask of object instead of rectangle include object. This modification make the … Continue Reading ››
  

• The importance of the balance of the training set and the times of each image used 【camera 151 ( left down )】
  1. each image is used for 15 to 40 times, but succeed about 11 times. 【statistical results】【result on test datasets】
     • 
  2. each image is used for 11 … Continue Reading ››
     

1. states of different cameras and corresponding training result (coco evaluation method)
   • camera 150 (right up): background color will change with my (people) different pose and different illumination (result on test datasets)
     • 
   • camera 151 (left down): background color will change with different illumination (result on test datasets)
     • Continue Reading ››
       

• QT-Net:
  1. Check the stability of the network. From the result, I think the network is stable.
  2. Optimize object extract, which lead to better combination result. But some bad combination still exist.
  3. Make a comparison on datasets whether to delete items which look very similar.
• SLAM:
  1. About the idea that I try about two week ago, I didn't find … Continue Reading ››
     

• Creating training and validation sets with images from more complex scenes, such as different illumination, different camera.
• Optimize the combination of objects to make all objects distribution more concentrated and reduce larger areas' occlusion.

In stereo matching, larger baseline will lead to better result. Also, some other SLAM methods perform well by setting keyframes with larger distance than adjacent frames. Therefore, I design an experiment to estimate the depth and pose from two non-adjacent frames ([0, 1], [0, 2], [0, 3]). The result shows that the depth … Continue Reading ››

Due to monocular pose estimation without scale information, I use several method to calculate the scale, and evaluate the pose result.

Make a analysis about the pose result from the self-supervised depth estimation. But with different evaluation indicators, the results are different. Maybe I make some mistakes.