Introduction

​

PT_network is a framwrok for deep learning experiments base on pytorch， which implements the following algorithms:

YOLOF:You Only Look One-level Feature(CVPR'2021)

FCOS: Fully Convolutional One-Stage Object Detection(ICCV'2019)

DataSets

The dataset type has VOCDataset and DetectionDataset.

VOCDataset

Follow the VOC dataset format:

├─Annotations
├─ImageSets
│  └─Main
└─JPEGImages

The json file is configured as follows:

    "dataset": {
        "input_size": [640, 640],
        "type": "VOCDataset",
        "train_data_path": "~/VOCdevkit/VOC2012",
        "val_data_path": "~/VOCdevkit/VOC2012"
    }

DetectionDataset

The annotations for the DetectionDataset is a json file generated with Labelme.

Images path and annotations path format is as follows:

├─train
│  ├─images
│  ├─annotations
└─val
    ├─annotations
    └─images

The json file is configured as follows:

"dataset": {
    "input_size": [640, 640],
    "type": "DetectionDataset",
    "train_data_path": "~/minidataset/train",
    "val_data_path": "~/minidataset/val"
},

Train

you can set parameters in "tools/config/det/yolof/yolof_resnet18_voc.json"

python /tools/train.py --config=tools/config/det/yolof/yolof_resnet18_voc.json

Result

Model	Backbone	Neck	FLOPS	DataSet(Close source)	AP
YOLOF	resnet18	DilatedEncoder	44.1	Industrial Dataset	0.883
FCOS	resnet18	NASFCOS_FPN	98.7	Industrial Dataset	0.8595
FCOS	resnet18	FPN	80.2	Industrial Dataset	0.9097

Special Thanks

Detectron2

MMDetection

YOLOF

FCOS