SSD

Why we included this project

This PyTorch code base reimplements SSD (Single Shot MultiBox Detector) with the pieces kept deliberately separate: the backbone, detector, box head, and predictor are all abstracted, so swapping in something like EfficientNet means registering it in a config file rather than rewiring the training loop. Multi-GPU training and batched inference work as expected, and checkpoints save the full state of the model, optimizer, and scheduler, so you can stop a run and resume exactly where you left off. TensorBoard metrics track per-category AP on VOC and the standard AP set on COCO while training runs. Inference also works on CPU, which makes this a practical base for teams that want to ship a detector without GPU servers at every stage. Researchers tweaking detector architectures will get the most out of it, but anyone who wants a readable SSD baseline to build on will find the structure easy to follow.