dicksonyuan / centernet-better

An easy to understand and better performance version of CenterNet

CenterNet

This repo is implemented based on my dl_lib, some parts of code in my dl_lib is based on detectron2.

Motivation

Objects as Points is one of my favorite paper in object detection area. However, its code is a little difficult to understand. I believe that CenterNet could get higher pts and implemented in a more elegant way, so I write this repo.

Performance

This repo use less training time to get a better performance, it nearly spend half training time and get 1~2 pts higher mAP compared with the old repo. Here is the table of performance.

Backbone	mAP	FPS	trained model
ResNet-18	29.8	92	google drive
ResNet-50	34.9	57	google drive
ResNet-101	36.8	43	google drive

What's New?

• treat config as a object. You could run your config file and check the config value, which is really helpful for debug.
• Common training / testing scripts in default. you just need to invoke dl_train/test --num-gpus x in your playground and your projects only need to include all project-specific configs and network modules.
• Performance report is dumped automaticly. After your training is over, we will evaluate your model automatically and generate a markdown file.
• Vectorize some operations. This improves the speed and efficiency.

What's comming

• Support DLA backbone
• Support Hourglass backbone
• Support KeyPoints dataset

Get started

Requirements

• Python >= 3.6
• PyTorch >= 1.3
• torchvision that matches the PyTorch installation.
• OpenCV
• pycocotools

pip install cython; pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

• GCC >= 4.9

gcc --version

Installation

Make sure that your get at least one gpu when compiled. Run:

pip install -e .

Training

For example, if you want to train CenterNet with resnet-18 backbone, run:

cd playground/centernet.res18.coco.512size
dl_train --num-gpus 8

After training process, a README.md file will be generated automatically and this file will report your model's performance.

NOTE: For ResNet-18 and ResNet-50 backbone, we suppose your machine has over 150GB Memory for training. If your memory is not enough, please change NUM_WORKER (in config.py) to a smaller value.

Testing and Evaluation

dl_test --num-gpus 8 

test downloaded model:

dl_test --num-gpus 8  MODEL.WEIGHTS path/to/your/save_dir/ckpt.pth 

Acknowledgement

• detectron2
• CenterNet

Coding style

please refer to google python coding style

AD time

I am an intern@megvii now, if you are interested in this company/want to get hired/want to be a intern, please send email to [email protected] or [email protected].