A faithful PyTorch port of the original SSD detector

What it does This is a from-scratch PyTorch implementation of the 2016 Single Shot MultiBox Detector. It trains and evaluates on PASCAL VOC and MS COCO, ships with pre-trained SSD300 checkpoints, and includes a live webcam demo for real-time object detection.

The interesting bit The port doesn’t just translate weights: training from scratch with data augmentation reaches 77.43% mAP on VOC2007, nudging past the original paper’s 77.2%. The dataset loaders also inherit from torch.utils.data.Dataset, which means they plug directly into the standard torchvision pipeline instead of inventing a new interface.

Key highlights

• Reproduces the original SSD300 results, hitting 77.43% mAP on VOC2007 with augmentation.
• Runs inference at roughly 45 FPS on a GTX 1060.
• Dataset loaders inherit from torch.utils.data.Dataset and slot into the torchvision pipeline.
• Ships with a live webcam demo that works on CPU or GPU.
• Offers both converted original Caffe weights and native PyTorch-trained checkpoints.

Caveats

• SSD512 and custom dataset training are still on the TODO list and not yet supported.
• The authors describe the project as a hobby, so maintenance is best-effort with “no guarantees.”

Verdict Grab this if you need a validated, classic SSD baseline in PyTorch for standard datasets or quick live demos. Skip it if you require SSD512, custom training pipelines, or an actively maintained framework.