search

mic-rud / GuidedDecoding

Accompaning repository for the 2022 ICRA paper "Lightweight Monocular Depth Estimation through Guided Decoding"
43 stars 10 forks source link

Encoder using Mobilnet V2 #4

Closed JiSuanJiDaWang closed 2 years ago

JiSuanJiDaWang commented 2 years ago

Hi, really impressive work! I want to use the MobilenetV2 as the encoder in your network, but the output of the mobilenet does not quit match the input of your encoder. I saw on your paper you did experiment on the MobilenetV2. Did you use upsampling and change the number of channel before the decoder or did you use a different decoder with more GUB than that in your paper? Could you give me some information? Thanks

mic-rud commented 2 years ago

Thank you for your interest! We used MobileNetV2 in combination with 2 Upsample Layers as described in FastDepth (https://github.com/dwofk/fast-depth), resulting in a feature map of 1/8 h x 1/8 w x 6. After that, 3 GUBs were used as for the DDRNet Version.

If you decide to "shorten" MobileNet and use GUBs on an intermediate feature map of the backbone, I would be very happy if you let me know about the results!

JiSuanJiDaWang commented 2 years ago

Thanks!