Github of Sebastian, Patrick and Sergej for CudaVision Lab - Final Project.
https://www.ais.uni-bonn.de/WS2223/CudaLab/
To Train with Coco dataset download cocoapi from: - https://github.com/cocodataset/cocoapi Follow install from given github for python.
- Implement Accuracy correctly (mAcc)
- Implement mIoU -> Metric apparently correctly implemented, but training on COCO doesn't show much improvement. Bug? Or requires more training?
- City-scapes dataset sequencing
- Pytorch dataset implementation -> Sort and sequence
- 5 options for the ground truth (Since it's 5 frame sequences)
- Temporal Augmentation
- Get baseline results (Frame by frame)
- Get results sequenced manner
- ConvGRU, ConvLSTM instead of ConvRNN for the COCO
- Encoder-Decoder (Resnet & VGG-Like)
- Check "temporal regularization"
- (Optional) U-Net (DRU and SRU)
- Generate images and notebook for results