# Deep Learning for Computer Vision ## from image classification to actual computer vision **ISAE-SUPAERO, SDD, 18 Nov. 2025** Florient CHOUTEAU <!--v--> Slides : https://fchouteau.github.io/isae-intro-to-cnns Notebooks : https://github.com/SupaeroDataScience/deep-learning/tree/main/vision <!--s--> # Use cases of CNNs <!--v--> We've done image classification that we applied in a sliding window fashion on larger images  <!--v--> We can solve other types of tasks with ConvNets  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> Keypoint Detection  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> 3D (VGGT) <video data-autoplay src="https://vgg-t.github.io/resources/teaser_video_v3_compressed_short.mp4"></video> <!-- .element: style="width: 60%; height: 40%"--> <!--v--> Image/Video Restoration-Upsampling  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> Image/Video Restoration-Upsampling  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> And of course image and video generation <video data-autoplay src="https://openaiassets.blob.core.windows.net/$web/nf2/blog-final/golden/6eda9a57-5d6d-4890-90ee-61f89e999719/20250922_1557_an%20astronaut%20golden%20retriever%20named%20_Sora_%20levitates%20around%20an%20intergalactic%20pup-themed%20space%20statio_simple_compose_01k5sta9hmegr9axrpbty5sxva%20(1).mp4"></video> <!--s--> # How ? It's just about architectures and loss functions <!--v--> Image Segmentation  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> Object Detection (intuition)  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> Object Detection (in practice)  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> Instance Segmentation  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> To learn more about this, see [this cs231n class](http://cs231n.stanford.edu/slides/2022/lecture_9_jiajun.pdf) <!--s--> # Encoder Decoder and U-nets <!--v--> How to solve a dense task with CNNs ? The problem : Context  <!-- .element: style="width: 60%; height: 40%"--> <!--v--> The Encoder Decoder Concept  <!--v--> ) <!-- .element: style="width: 60%; height: 40%"--> <!--v--> How to upsample ? Static upsampling  <!-- .element: style="width: 50%; height: 50%"--> <!--v--> How to upsample ? Learnable upsampling  <!-- .element: style="width: 50%; height: 50%"--> <!--v--> How to upsample ? Learnable upsampling  <!-- .element: style="width: 50%; height: 50%"--> <!--v--> Is it enough ? How to keep spatial information ?  <!-- .element: style="width: 50%; height: 50%"--> <!--v--> They are called [U-Net (2015)](https://arxiv.org/abs/1505.04597)s  <!-- .element: style="width: 50%; height: 50%"--> <!--v--> Imagen & co are based on unets  <!--s--> # Hands-on 2 : From image classification to image segmentation <!--v--> ### Dataset description - 12800 train images (6400 cloudy, 6400 clear), size 64x64 - 640 test images (320 cloudy, 320 clear), size 64x64  <!-- .element height="40%" width="40%" --> <!--v--> ### Dataset description - 64 big test images (256x256)  <!-- .element height="40%" width="40%" --> <!--v--> ### Let's go ! - 3_classification_to_segmentation.ipynb : From image classification to dense tasks - 4_sliding_window_unet.ipynb : How to apply our detector to a big satellite image ? - <!--s--> Extra : Artificial Intelligence in practice in a space systems context [ISAE SPAPS 2023-2024 class](https://docs.google.com/presentation/d/10zd65eg0X_aqydggKRvC3s20AXSu3WI9RWWUKJVvfAI/edit?usp=sharing) <!--s--> Extra : Other keywords that are important for CNN in Computer Vision - [Self-Supervised Learning](http://cs231n.stanford.edu/slides/2023/lecture_13.pdf) - [Vision Transformers](https://www.v7labs.com/blog/vision-transformer-guide) - [Generative Adversarial Networks](https://github.com/SupaeroDataScience/deep-learning/tree/main/GAN) - [Diffusion Models](http://cs231n.stanford.edu/slides/2023/lecture_15.pdf) (StableDiffusion, MidJourney ...)
