-
Implement Bayesian SegNet for semantic segmentation with Pytorch, and also generate estimates of aleatoric and epistemic uncertainties associated with the segmentation. Relevant codes and files are stored in
./Bayesian, -
Implement UNet for semantic segmentation with TensorFlow. Stop training due to a lack of GPUs. Relevant codes and files are stored in
./UNet.For more information on the dataset please refer to: CityScapes dataset. This is also a team project for Deep Learning for Computer Vision course, lectured by Prof. Alexander Amini.
git clone git@github.com:glisses/segmentation-and-uncertainty-estimation-on-CityScapes.git
pip install -r requirements.txt
Downloaded vgg16_bn-6c64b313.pth from https://download.pytorch.org/models/vgg16_bn-6c64b313.pth and put it in the same folder as './BayesianSegNet/'.
Put lab2_train_data.h5 and lab2_test_data.h5 in './BayesianSegNet/'.
-
./BayesianSegNet/main_segnet_v7.ipynbis the main file of this software lab. -
Ignore the first 4 cells in
./BayesianSegNet/main_segnet_v7.ipynbif you're not using Google Colab. -
To train the model, change the parameter
MODEto'TRAIN'and run all cells in./BayesianSegNet/main_segnet_v7.ipynb. -
To test the model only, change the parameter
MODEtoTESTand run all cells in./BayesianSegNet/main_segnet_v7.ipynb. Model parameters are stored in./BayesianSegNet/weights/23_model.pth.
| Team Member | Contribution |
|---|---|
| Jialong Yuan | Finish the exercises and implement a data loader. |
| Zhuoyuan Li | Provide code of the UNet model structure using TensorFlow. |
| Yitong Luo | Implement UNet for segmentation, including training and testing. |
| Kaiang Wen | Assign tasks; Implement Bayesian SegNet for segmentation; Generate and visualize estimates of aleatoric and epistemic uncertainties with Yaqi Zhou. |
| Yaqi Zhou | Generate and visualize estimates of aleatoric and epistemic uncertainties; Help train UNet and write comments for it. |
| ALL | Check through the team project. |