AI-Related Notes
Common vision models and their characteristics, organized by image recognition, object detection, image segmentation, pose estimation, and generation tasks.
Updated
Existing Vision Models
| Domain | Model | Description |
|---|---|---|
| Convolutional neural networks | LeNet | An early convolutional network, used for handwritten digit recognition |
| Convolutional neural networks | AlexNet | The landmark model that marked deep learning's breakthrough in image recognition |
| Convolutional neural networks | VGGNet | Builds deep networks by repeatedly using simple convolutional and pooling layers |
| Convolutional neural networks | ResNet | Introduces residual connections to solve the difficulty of training deep networks, allowing deeper models |
| Convolutional neural networks | Inception (GoogLeNet) | Increases network width and depth through a "network in network" structure |
| Object detection | Faster R-CNN | Effectively combines a region proposal network (RPN) with Fast R-CNN for object detection |
| Object detection | YOLO | Treats the task as a single regression problem, achieving fast object detection |
| Object detection | SSD | Directly predicts bounding boxes and class probabilities at multiple scales, balancing speed and accuracy |
| Image segmentation | FCN (Fully Convolutional Network) | Used for semantic segmentation, classifying every pixel in an image |
| Image segmentation | UNet | Especially suitable for medical image segmentation, with a distinctive U-shaped structure |
| Image segmentation | Mask R-CNN | Adds a branch on top of Faster R-CNN to generate high-quality segmentation masks |
| Pose estimation | OpenPose | A real-time multi-person system that can estimate poses from images |
| Generative models | VAE (Variational Autoencoder) | Learns a latent representation of given data through an encoder and a decoder |
| Generative models | GAN (Generative Adversarial Network) | Consists of a generator and a discriminator, used to generate realistic images |
| Generative models | CGAN (Conditional Generative Adversarial Network) | Adds a conditional variable on top of GAN, allowing the model to generate images with specific features |