MMDetection rebuilt for the TensorRT escape route

What it does vedadet is a PyTorch toolbox for training single-stage object detectors such as RetinaNet and FCOS. It splits the detection pipeline into four modular stages—data pipeline, model, postprocessing, and criterion—with the explicit goal of making PyTorch models portable to TensorRT engines on NVIDIA devices. The codebase is heavily derived from mmcv and mmdetection, but refactored around deployment rather than research agility.

The interesting bit Most detection frameworks treat model conversion as a postscript; vedadet treats it as the architectural spine. By isolating postprocessing and criterion from the model itself, it tries to ensure the path from training script to Jetson Nano is mostly frictionless.

Key highlights

• Ships with RetinaNet, FCOS, and other single-stage detectors ready to train.
• Four-way modularity: data, model, postprocessing, and criterion are intentionally decoupled.
• Explicitly targets NVIDIA hardware including Tesla V100, Jetson Nano, and Jetson AGX Xavier.
• Exports to ONNX with static or dynamic input shapes.
• Pairs with dedicated Python (flexinfer) and C++ (cheetahinfer) inference front-ends.

Caveats

• Group Normalization and Deformable Convolution are unsupported for ONNX export.
• Linux and CUDA 10.2+ are mandatory; the documented testbed uses Ubuntu 16.04 and PyTorch 1.6.0.

Verdict Grab this if you need single-stage detectors that must eventually live on NVIDIA edge hardware and you want the training-to-deployment pipeline to stay in one family. Skip it if your work depends on deformable convolutions or you are looking for a general-purpose, OS-agnostic research framework.