Segment Anything: A Promptable Foundation Model for Image Masks
Kirillov et al. built a ViT image encoder, a prompt encoder, and a mask decoder trained on a billion masks. Why promptable segmentation is a product interface, not just a COCO score.
Tag
14 articles · 39 min of reading
Kirillov et al. built a ViT image encoder, a prompt encoder, and a mask decoder trained on a billion masks. Why promptable segmentation is a product interface, not just a COCO score.
Radford et al. trained image and text encoders to match captions at web scale. Zero-shot classifiers, retrieval, and why the prompt template is part of the model.
Krizhevsky, Sutskever, and Hinton trained a deep ReLU CNN on two GPUs and crushed ILSVRC 2012. Dropout, data augmentation, and the moment vision switched to learned features.
Tan and Le replaced one-dimensional scaling with a compound coefficient that grows depth, width, and input size in a fixed ratio. NAS-found mobile blocks, and when compound scaling is the wrong knob.
Ren et al. added a Region Proposal Network that shares features with Fast R-CNN so two-stage detection could run in near real time. Anchors, RoI pooling, and when two stages still beat one shot.
Deng et al. built a WordNet-organized image dataset at a scale that made large-scale classification a community sport. Why the benchmark still shapes models, and how its biases leaked into production.
Szegedy et al. (GoogLeNet) stacked 1×1, 3×3, and 5×5 paths with pooling, using 1×1 bottlenecks to keep FLOPs in check. Auxiliary classifiers, and why multi-scale blocks still reappear.
He et al. mask a large fraction of patches, encode the visible few, and decode pixels with a light decoder. Why 75% masking is the point, and how MAE pretraining changed vision self-supervision.
Chen et al. showed that a simple InfoNCE setup with strong augmentations, a projection head, and large batches learns ImageNet features without labels. What to copy, and why the batch is part of the algorithm.
Ronneberger et al. built an encoder-decoder with cropped skips so biomedical images could be segmented from few annotated slides. Why the U-shape still shows up in diffusion and medical stacks.
Dosovitskiy et al. dropped the conv inductive bias, split images into patches, and trained a transformer on ImageNet-21k-scale data. When ViTs beat ResNets, and when they need the extra data or distillation.
Redmon et al. framed object detection as a single CNN pass that predicts boxes and class probabilities on a spatial grid. Speed as a product constraint, and the localization trade that later YOLO versions kept renegotiating.
He et al. 2016 showed degradation was an optimization problem. Residual blocks, identity shortcuts, and what that still means for modern networks.
How Pinterest built visual search and object detection to let people discover products and ideas by tapping on specific items inside an image.