pierluigiferrari/ssd_keras

Implements SSD300, SSD512, and a lightweight SSD7 variant using VGG-16 and custom base networks, with multi-scale feature maps for detecting objects at various sizes in a single forward pass. Provides ported weights matching original Caffe implementation mAP (77.5-83.2% on Pascal VOC) plus transfer learning utilities for custom datasets. Emphasizes documentation and code clarity with detailed comments, Jupyter tutorials, and template architectures to facilitate model customization and understanding.

1,873 stars. No commits in the last 6 months.