DeepLabV3 with MobileNetV2

This is the pytorch implementation of DeepLabV3 segmentation with MobileNetV2 support backbone. It achieve STOA speed and meanIOU on semantic segmentation. Benefit from MobileNetV2 depth-wise convolution and DeepLabV3 the most advanced ASPP module, the segmentation result is remarkable. Here is some screen shot of result:

This is only about 23 epoch result, further result maybe update later. For now, DeepLabV3 with MobileNetV2 has those features and you can not reject it:

• Fast: almost 25 fps on GTX1080, it's almost 80% faster than original DeeplabV3;

• Accurate: compare to ENet or SegNet or UNet or RetinaSeg, it achieve almost 78 meanIOU on test dataset;

• Without post process with good result, as you can see, the result can almost use without CRF post process.

Install

To run:

sudo pip3 install alfred-py
python3 demo.py

Further training

For training, you can obtain full version from StrangeAI

Contact

Any question could be asked via Setu(a secret chat app): http://loliloli.pro