Computer Vision Lab Model cards
Model cards
47 results
- MobileNetV2Inverted residuals + linear bottlenecksA lightweight classifier officially trained in torchvision, the default model in the lab. It also outputs the last feature maps to compute Grad-CAM, and the map…Runs in the browserArchitecture diagram
- MobileNetV3-LargeNAS-designed MobileNet with SE and h-swishA newer version of MobileNet designed with automated architecture search, with Squeeze-and-Excitation units.Runs in the browserArchitecture diagram
- EfficientNet-B0Compound-scaled MBConv networkThe smallest of the EfficientNet family, which scales depth, width and resolution with a compound coefficient. The lab uses bilinear interpolation in preprocessing instead of the…Runs in the browserArchitecture diagram
- ResNet-18Residual network, 18 layersThe smallest ResNet model, a direct example of residual connections.Runs in the browserArchitecture diagram
- ResNet-50 (IMAGENET1K_V2)Residual network with bottleneck blocksA reference model widely used in research, with the improved torchvision weights (V2).Runs in the browserArchitecture diagram
- DeiT-Tiny (ViT-Ti/16)Vision Transformer, 12 layers × 3 heads, patch 16A small vision transformer trained on ImageNet alone. It outputs the attention matrices of all layers, so it is used for Attention Rollout. The lab uses…Runs in the browserArchitecture diagram
- SSDlite320 MobileNetV3-LargeSingle-shot detector with depthwise headsA lightweight detector that runs in the browser. Box decoding and NMS are done in JavaScript with the original model's anchors.Runs in the browserArchitecture diagram
- LR-ASPP MobileNetV3-LargeLite R-ASPP segmentation headA lightweight semantic segmentation model with 21 classes (Pascal VOC).Runs in the browserArchitecture diagram
- DeepLabV3 MobileNetV3-LargeAtrous convolution + ASPPSemantic segmentation that is more accurate and heavier than LR-ASPP, for comparing the two models.Runs in the browserArchitecture diagram
- MONAI Ventricular Short-Axis U-Net2D U-Net, 3 labels + backgroundA U-Net model for segmenting the cavities of the two ventricles and the heart muscle in 256×256 short-axis cardiac MRI slices. For educational purposes only, and…Runs in the browserArchitecture diagram
- Faster R-CNN ResNet-50 FPNTwo-stage detector with RPN + FPNA two-stage detector that runs only through the optional Python server.Requires the optional Python serverArchitecture diagram
- Mask R-CNN ResNet-50 FPNFaster R-CNN + mask headReal instance segmentation, through the optional server.Requires the optional Python serverArchitecture diagram
- ViT-B/16Vision Transformer Base, patch 16Knowledge card: the original vision transformer in its base size, too large to run in the browser.Knowledge cardArchitecture diagram
- VGG-16Stacked 3×3 convolutionsKnowledge card: a simple, deep architecture with many parameters.Knowledge cardArchitecture diagram
- AlexNet5 conv + 3 FC layersKnowledge card: the model that launched the wave of deep learning in vision in 2012.Knowledge cardArchitecture diagram
- LeNet-52 conv + 3 FC layersKnowledge card: one of the first successful convolutional networks, for reading handwritten digits.Knowledge cardArchitecture diagram
- The YOLO familyOne-stage grid-based detectorsKnowledge card: fast one-stage detectors. No version of it is loaded in the lab because the licences differ between versions; check the version's licence before using…Knowledge cardArchitecture diagram
- U-Net (original)Encoder–decoder with skip connectionsKnowledge card: the original architecture for biomedical segmentation.Knowledge cardArchitecture diagram
- nnU-NetSelf-configuring U-Net frameworkKnowledge card: a framework that automatically configures the preprocessing, architecture and training for each dataset, and a strong reference in medical competitions.Knowledge cardArchitecture diagram
- SegFormerHierarchical transformer encoder + MLP decoderKnowledge card: semantic segmentation with a hierarchical transformer. Not loaded because the licence of the official weights does not allow distribution.Knowledge cardArchitecture diagram
- FCN ResNet-50Fully convolutional networkKnowledge card: the first fully convolutional network for semantic segmentation.Knowledge cardArchitecture diagram
- GoogLeNet (Inception v1)Inception modules with 1×1 dimension reductionKnowledge card: the winner of the ILSVRC 2014 classification challenge; Inception modules combine filters of different sizes in parallel and use 1×1 convolutions for dimensionality reduction,…Knowledge cardArchitecture diagram
- Inception-v3Factorised Inception modules, 299×299 inputKnowledge card: a development of GoogLeNet that factorises large convolutions into small or asymmetric ones (1×n then n×1), reduces the grid with parallel branches, and uses…Knowledge cardArchitecture diagram
- DenseNet-121Densely connected convolutional network, growth rate 32Knowledge card: each layer inside a dense block receives the feature maps of all the previous layers by concatenation, so features are reused and gradient flow…Knowledge cardArchitecture diagram
