2023-Winter-Frontier-Research-Program

Deep Learning in Computer Vision

If you find my project useful, don't hesitate to give it a star.

Part 1: Digit recognition on MNIST

Architecture of my model:

• [conv-relu-pool]x3 -> [affine]x1 -> softmax

1. Convolutional layer (with bias) with 24 5x5 filters, with zero-padding of 2, with stride of 1
2. ReLU
3. pooling layer with window size of 2x2, with stride of 2, without padding
4. Convolutional layer (with bias) with 48 5x5 filters, with zero-padding of 2, with stride of 1
5. ReLU
6. pooling layer with window size of 2x2, with stride of 2, without padding
7. Convolutional layer (with bias) with 64 5x5 filters, with zero-padding of 2, with stride of 1
8. ReLU
9. pooling layer with window size of 2x2, with stride of 2, without padding
10. Fully-connected layer (with bias) to compute scores for 10 classes

Training and validation accuracy:

Test: 99.15%

Part 2: Semantic Segmentation and Uncertainty Estimation on CityScapes

• Used PyTorch to build a semantic segmentation pipeline for autonomous driving based on U-Net.

• Estimated aleatoric and epistemic uncertainty in my model and its predictions. Key Idea: Do dropout at both training and testing time. At test time, repeat prediction a few hundreds times with random dropout. The variance of predictions gives the episdemic uncertainty of the model.
  Reference: Uncertainty in Deep Learning. How To Measure?

  After adding dropout layers: Bayesian U-Net

  

  Results