How residual networks use skip connections to train very deep CNNs for image classification.

The region-based CNN family for object detection and segmentation.

Berkeley's conditional GAN for paired image-to-image translation.

The encoder–decoder architecture that became a standard for biomedical and general segmentation.

Oxford's deep, uniform 3×3 convolutional architecture for large-scale recognition.

The network that reignited deep learning by winning ImageNet 2012.

Reconstructing 3D head models from unconstrained internet photos across pose and lighting.

Predicting how memorable an image is, inspired by human cognition.

Reasoning about relations between entities over time for video understanding.

Skull-based facial reconstruction for forensic science and anthropology.

Learning realistic garment wrinkles for body shape and pose estimation.

A content loss that preserves style for high-resolution, real-time style transfer.