realtime_object_detection's Introduction

Realtime_Object_Detection

Codebase for realtime object detection

News: With the help of this codebase, we win the second place in the 2019 DAC System Design Contest.

Introduction

This codebase is modified based on mmdetection and the difference between them is that some two stage object detection algorithms are removed and some small and compact backbones are added.

The purpose of this codebase is to facilitate network quantization and pruning, eventually applying deployment on different hardware devices.

Installation

Requirements

Linux
Python 3.5+
CUDA 9.0+
NCCL 2+
GCC 4.9+

Step by step installation

a. Create a conda virtual environment and activate it.

conda create -n realtime_object_detection python=3.7 -y
source activate realtime_object_detection

b. Install some requirements.

pip install torch torchvision
pip install cython

c. Clone the realtime_object_detection repository.

git clone https://github.com/waterbearbee/realtime_object_detection
cd realtime_object_detection

d. Compile cuda extensions.

sh ./compile.sh

e. Install other dependencies.

python3 setup.py develop

Getting Started

Training

Train with a single GPU

python3 ./tools/train.py ${CONFIG_FILE}

Train with multiple GPUs

python3 -m torch.distributed.launch --nproc_per_node=${GPU_NUM} ./tools/train.py ${CONFIG_FILE} --launcher pytorch

Testing

Test with a single GPU

python3 ./tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] --eval bbox

Test with multiple GPUs

python3 -m torch.distributed.launch --nproc_per_node=${GPU_NUM} ./tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --out ${RESULT_FILE} --eval bbox --launcher pytorch

DAC Competition Results (RetinaNet)

Backbone	Neck	Head	box AP
MobileNetv2_x1_0	P3-P5	2conv(64c)	84.2
MobileNetv2_x0_5	P3-P5	2conv(64c)	79.9
MobileNetv2_x0_25	P3-P5	2conv(64c)	74.4
ShuffleNetV2_x1_0	P3-P5	2conv(64c)	80.6
ShuffleNetV2_x0_5	P3-P5	2conv(64c)	75.0

Recommend Projects