Implementation of DL architectures for semantic segmentation of images and point clouds: vanilla U-Net [1], DeepLabV3 [2], DeepLabV3+ [3], RandLaNet [4]
Train scripts ./segmentation_models/train_vaihingen.py and ./segmentation_models/ train_dales.py for Vaihingen and Dales datasets.
Inference scripts ./inference/inference_vaihingen.py and ./inference/inference_dales.py
- DaLes Dataset semantic segmentation with RandLaNet-5
excellent:
and bad:
here blue - GLO, green - vegetation, red - buildings, magenta - clutter.
Metrics: 0.80 mIoU, GLO 0.84 IoU, vegetation 0.79 IoU, buildings 0.75 IoU, clutter 0.75 IoU.
- Buildings detection (one class segmentation) for Massachusetts datasets with DeepLabV3+ (ResNet-152 backbone) 0.85 IoU
- Real life aerial pics segmentation examples:
Trained models RandLaNet and DeepLabV3+ models
[1] Olaf Ronneberger and Philipp Fischer and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation
[2] Liang-Chieh Chen and George Papandreou and Florian Schroff and Hartwig Adam. Rethinking Atrous Convolution for Semantic Image Segmentation
[3] Liang-Chieh Chen and Yukun Zhu and George Papandreou and Florian Schroff and Hartwig Adam. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation
[4]Qingyong Hu and Bo Yang and Linhai Xie and Stefano Rosa and Yulan Guo and Zhihua Wang and Niki Trigoni and Andrew Markham. RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point Clouds