ENS Paris-Saclay M1 IA&ML: Image and video - 2023

Gabriele Facciolo

General info

ENS Paris-Saclay

Each course will have a part of Lecture and a part of Practical Session. The Practical Session will be based on Keras and Pytorch using colab notebooks.

Please, make sure to bring your own laptop

Schedule

1. – Monday 18/9 (1X74): 8h00-12h30 - Intro NN, backprop and CNN for classification

2. – Friday 22/9 (1Q07): 13h30-16h00 - Semantic segmentation

3. – Friday 29/9 (1Q07): 13h30-16h00 - Object detection

4. – Friday 6/10 (1Q07): 13h30-16h00 - Transfer learning and representation learning

Slides and notebooks

1. Slides 1

   Notebook 0: jupyter notebook tutorial

   Notebook 0.5: not pytorch tutorial

   Notebook 1: Classification CNNs

2. Slides 2

   Notebook 2: Semantic segmentation

   Notebook 2.5: More semantic segmentation with pytorch

3. Slides 3

   Notebook 3.1: YoloV1 vehicle detection (broken)

   Notebook 3.2: YoloV2 Step-by-Step

   Notebook 3.3: Mask R-CNN semantic segmentation with pytorch

   Notebook 3.3: Mask R-CNN semantic segmentation with keras

   Notebook 3.4: Human pose estimation

4. Slides 4

   Notebook 4.1: Transfer learning with keras

   Notebook 4.2: Interacting with CLIP

   Notebook 4.2.5: Training CLIP (much more details)

Evaluation: mini projects

The idea is to write a report in the form of an article (~10 pages) describing a method. Plus a notebook that would allow to reproduce some experiments (but not the training). Plus a 15 min presentation (with questions included). The structure of the “article” could be:

• Introduction: description of the problem and main ideas of the current paper.
• Method and Architecture: detailed description of the method
• Data: Description of the data used for training
• Training: explain how the network is trained
• Inference: explain how the network is uses at inference time
• Experiments: present some results
• Limitations: show limitations

subjects: to be determined