Advanced-Computer-Vision

Computer Vision projects

Content:

5. Batch normalization layer and residual connections

   • Training degradation of very deep CNNs
   • Countermeasure: Batch normalization
   • Countermeasure: Skip-connections

6. Saliency analysis

   • Model: ResNet50 pretrained on ImageNet dataset
   • Saliency map
   • Receptive fields

7. Metric learning

   • Dataset: MNITS
   • Embedding space by classification task
   • Embedding space by contrastive task

8. Image Segmentation

   • Dataset: CWFID (crops vs weeds)
   • Segmentation with U-net

9. Transfer learning

   • Pre-training dataset: Deep Weeds
   • Pre-training task: multiclass classification
   • Fine-tuning dataset: CWFID
   • Fine-tuning task: semantic segmentation with U-net

10. Adversarial attacks

    • Carlini-Wagner Attack on MNIST classification model
    • Sparse pertubation with Hoyer-Square regularizer