Single Shot Detection

Build flexible object detection pipelines with declarative configuration

Content

You are being provided with the following set of features:

• Supporting latest PyTorch release
• Train SSD, M2Det, RetinaNet or some custom architecture
• Available backbones: torchvision + pretrainedmodels + custom MobileNet, MobileNetV2
• Data augmentations
• AdamW and SGDW optimizers, some custom learning rate schedulers
• Weight pruning for efficient inference
• Export to ONNX or OpenVINO
• Tensorboard integration
• Training callbacks

Quick start

Download PASCAL VOC or COCO dataset and start training using one of the provided sample configs:

python3 main.py --config samples/ssd_mb2_voc.py

^{(don't forget to adjust paths in the config first!)}

To see the list of parameters:

python3 main.py --help

Requirements

• python 3.6
• opencv with python bindings
• libturbojpeg
• requirements.txt

Structure

Some places that may be useful to look into:

• bf - provides common reusable parts for building a deep learning pipeline
  • bf.base - custom backbone network implementation (e.g. MobileNetV2)
  • bf.datasets - dataset handling
  • bf.preprocessing - data augmentations and preprocessing
  • bf.training - callbacks; custom optimizers and learning rate schedulers; weight prunner
  • ...
• detection - parts of code which are used to build object detection pipelines on top of bf
• samples - contains sample configuration files for popular network architectures
• main.py - the entry point

Inspired by

• Tensorflow Object Detection API (https://github.com/tensorflow/models/tree/master/research/object_detection)
• Gluon CV (https://github.com/dmlc/gluon-cv)
• mmdetection (https://github.com/open-mmlab/mmdetection)