• Using the light weight MobileNet as the encoder for depth predicting.
• In the decoder, using depthwise and pointwise convolution replace the traditional convolution.
• The result shows that the light weight network can achieve comparable accuracy via the instruction of a strong network.

２月９日：１０：１０　E7-414 　リハーサル

二石 航歩、水戸竣哉、財前　智行、五十君良寿

２月１０日：１０：００　E7-414

Du Xueting、Gao Yuliang　リハーサル

２月１４日：９：００　E7-414

全員仕上げ練習。

• Self-supervised training monocular depth prediction network under the instruction of a stronger network that trained on stereo image. The result shows that the accuracy can be improved, and the scale can be recovered.
• About the question: network perform well on classification task while has no advantage in depth prediction task. I read some papers … Continue Reading ››
  

1. QT-Net:
   • Finished the creating of docker image.
   • Transferred the pytorch model to IR model, so that the network can be run faster on AE2100.
2. SLAM
   • Testing the influences of each part work to the accuracy of depth and pose estimation. Stronger backbone can lead to a higher accuracy. However, light-weight backbone which can lead to a … Continue Reading ››
     

1. QT-Net
   • Taking source image for another 50 classes of items.
   • To test the capacity of the network, doing experiments on the datasets with 90 classes.
   • Making annotation for the test datasets with 90 classes, and test the model on it, the accuracy is 0.987 (IOU=0.75)
   • Modified the output of the plot of BBox. For any item, … Continue Reading ››
     

1. QT-NET:
   • Take three group of data for another 58 kinds of items.
2. SLAM: do experiments about estimating depth from stereo images
   • Estimating depth by inputting left image and create loss function by reprojecting left image from right image. The comparable accuracy is achieved.
   • Estimating depth by inputting left and right image, and create cost volume by … Continue Reading ››
     

1. QT-NET:
   • Done a experiment on the 3 group source image, and the trained model can achieve 0.96 AP on the validation datasets.
   • About the testing, it can achieve high accuracy on the 512x512 image, cannot detect image from 600x800 image. And testing on one image takes about 0.06s.
2. SLAM:
   • Paper reading: undeepvo: monocular visual … Continue Reading ››
     

出席：全員

研究進捗確認、国際会議投稿確認、研究指示をおこなった。

Mr. GAO AROB論文投稿添削。