Awesome

yolo2_light

Light version of convolutional neural network Yolo v3 & v2 for objects detection with a minimum of dependencies (INT8-inference, BIT1-XNOR-inference)

This repository supports:

How to compile:

To compile for CPU just do make on Linux or build yolo_cpu.sln on Windows
To compile for GPU set flag GPU=1 in the Makefile on Linux or build yolo_gpu.sln on Windows

Required both CUDA >= 8.0 and cuDNN >= 7.1.1

How to start:

Download yolov3.weights to the bin directory and run ./yolo.sh on Linux (or yolo_cpu.cmd / yolo_gpu.cmd on Windows)
Download yolov3-tiny.weights to the bin directory and run ./tiny-yolo.sh

How to use INT8-inference:

Use flag -quantized at the end of command, for example, tiny-yolo-int8.sh or yolo_cpu_int8.cmd
For the custom dataset, you should use input_calibration= parameter in your cfg-file, from the correspon cfg-file: yolov3-tiny.cfg or yolov3.cfg, ...

How to use BIT1-XNOR-inference - only for custom models (you should train it by yourself):

You should base your cfg-file on yolov3-spp_xnor_obj.cfg and train it by using this repository as usual https://github.com/AlexeyAB/darknet by using pre-trained file darknet53_448_xnor.conv.74
Then use it for Detection-test or for getting Accuracy (mAP):
- ./darknet detector test data/obj.names yolov3-spp_xnor_obj.cfg data/yolov3-spp_xnor_obj_5000.weights -thresh 0.15 dog.jpg
- ./darknet detector map data/obj.data yolov3-spp_xnor_obj.cfg data/yolov3-spp_xnor_obj_5000.weights -thresh 0.15