Dog-O-Matic breed classifier

INT8 ONNX model for in-browser dog breed classification, used by Dog-O-Matic.

• Backbone: convnext_nano
• Classes: 127 breeds
• Input: 224x224 RGB, ImageNet normalization
• Validation: top-1 84.0%, top-5 97.5% (Tsinghua Dogs, balanced 40/breed)

Scope and limitations

This model reports visual resemblance to purebred classes, not ancestry. Mixed-breed dogs frequently look nothing like their actual parentage: a softmax over purebred classes is a confusion distribution, not an ancestry distribution, and should not be read as one.

Not valid for legal, housing, insurance, or breeding decisions, and not a source of health or veterinary information. Visual breed identification is demonstrably unreliable -- in the largest shelter study, staff assigned breeds absent from a dog's DNA entirely for a third of dogs.

Training data

Tsinghua Dogs (70,428 images, 130 breeds). Stanford Dogs is deliberately not used as a benchmark: it is derived from ImageNet, so ImageNet-pretrained backbones have seen its test images with labels.