mmdetection-annotated

Note on mmdetection for better usage and understanding.
Notice : Focused on some experiments recently , I may postpone updates on training parts.To be contunued....(well,I promise to finish the rest work before Jul.Blame it on examinations...)

Introduction

Refer to the execllent implemention here:https://github.com/open-mmlab/mmdetection ,and thanks to author Kai Chen.
Open-mmlab project , which contains various models and implementions of latest papers , achieves great results in detection/segmentataion tasks , and is kind enough for rookies in CV field.

Getting started

More information about installation or pre-train model downloads , pls refer to officia mmdetection or blog here

Test on images
You can test on Mask-RCNN demo by running the script demo.py. I have just rewritten the demo file to detect on single image or a folder as follow:

import ipdb
import sys,os,torch,mmcv
from mmcv.runner import load_checkpoint
#下面这句import的时候定位并调用Registry执行了五个模块的注册
from mmdet.models import build_detector	
from mmdet.apis import inference_detector, show_result

if __name__ == '__main__':
	# ipdb.set_trace()
	cfg = mmcv.Config.fromfile('configs/mask_rcnn_r101_fpn_1x.py')
	# cfg = mmcv.Config.fromfile('configs/faster_rcnn_r50_fpn_1x.py')
	cfg.model.pretrained = None		#inference不设置预训练模型
	#inference只传入cfg的model和test配置，其他的都是训练参数
	model = build_detector(cfg.model, test_cfg=cfg.test_cfg)
	_ = load_checkpoint(model, 'weights/mask_rcnn_r101_fpn_1x_20181129-34ad1961.pth')
	# print(model)  # 展开模型

	# test a single image
	img= mmcv.imread('/py/pic/2.jpg')
	result = inference_detector(model, img, cfg)
	show_result(img, result)

	# # test a list of folder
	# path='/py/mmdetection/images/'
	# imgs= os.listdir(path)
	# for i in range(len(imgs)):
	# 	imgs[i]=os.path.join(path,imgs[i])
	# # imgs = ['/py/pic/4.jpg', '/py/pic/5.jpg']
	# for i, result in enumerate(inference_detector(model, imgs, cfg, device='cuda:0')):
	#     print(i, imgs[i])
	#     show_result(imgs[i], result)

Debug
You can debug by setting breakpoint with method of adding ipdb.set_trace()
Hook
If you want to inspect on intermediate variables , hook.py can be a provision served as a reference for your work.

Annotations

Annotations are attached everywhere in the code(surely only the part I have read , and the not finished part will be completed as soon as possible). Beside , annotation folder contains some interpreting documents as well.

Model visualization
Take Mask-RCNN for example , the model can be visualized as follow:(more details refere to model-structure-png)

Configuration
Explicit describtion on config file , take Mask RCNN for example , refer to mask_rcnn_r101_fpn_1x.py
MMCV&MMDET
Specification of mmcv lib and a partial of mmdet(more details about various models will be updated later ).

Detection Results

Test on Mask RCNN model:

Training

dataset

You can just use COCO dataset , refer here.
If you want to train on your customed dataset labeled by labelme , you need first convert json files to COCO style , this toolbox may help you ;
If you want to train on your customed dataset labeled by labelImg , you need first convert xml files to COCO style , this toolbox may also help you .
I have tested on these tools recently to make sure them still work well, if questiones still arised , desrcibe on issue please or contact me , thanks.

learning rate

Remember to set lr in config file according to your own GPU_NUM !!!!(eg.1/8 of default lr for 1 GPU)

Future work

The training part is not finished yet , and continue work need to be done . Mmdetection performs better than many classical implementions , it's really a excellent work , can be called as ‘Chinese Detectron’ :p .I will update this project with annotations for training part , letting more people make a good use of this great work.You can continue to foucus on this repo.
BTW , this repo is just used for better comprehension , if you ask for better performance or latest paper implementions ,please keep eyes on mmdetection

Dwrety / mmdetection-annotated