Mask_RCNN_mmdet

Selected Topics in Visual Recognition using Deep Learning Homework 3 announcement HW3

This Rep is base on open-mmlab/mmdetection.

More detail can see mmdetection getting start.

Hardware

• Ubuntu 18.04 LTS
• Intel(R) Core(TM) i7-7820X CPU @ 3.60GHz
• 31 RAM
• NVIDIA GTX 1080 12G *1

ENV

• CUDA 10.2
• Pytorch 1.6.0
• mmcv-full latest+torch1.7.0+cu110
• mmdetection master(2020/12/11)

Install env

# getting start mmdectation
# https://github.com/open-mmlab/mmdetection/blob/master/docs/get_started.md

# Pytorch 1.6.0  CUDA 10.2
pip3 install torch==1.7.0+cu110 torchvision==0.8.1+cu110 torchaudio===0.7.0 -f https://download.pytorch.org/whl/torch_stable.html

# Install mmcv-full, we recommend you to install the pre-build package as below.
pip3 install mmcv-full==latest+torch1.7.0+cu110 -f https://download.openmmlab.com/mmcv/dist/index.html

# Clone the MMDetection repository.
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection

# Install build requirements and then install MMDetection.
pip3 install -r requirements/build.txt
pip3 install -v -e .  # or "python setup.py develop"

Reproducing Submission

To reproduct my submission without retrainig, do the following steps:

1. Dataset Preparation
2. Training
3. Inference

Dataset Preparation

Prepare Images

After downloading images and spliting json file, the data directory is structured as:

 data/
  | +- train_images/
  | +- test_images/
  | -- pascal_train10.json
  | -- pascal_train90.json
  | -- test.json

Download Classes Image

Dataset: https://drive.google.com/drive/folders/1nglaZBJJ_Amonndw4nIVBh_UuCpp4gee?usp=sharing

Download and extract train_images.zip and test_images.zip to data directory.

Splited training and validation json used coco style

Use split8020.py to make train.txt .

$ python3 split8020.py

Training

Setting

You can setting detail Hyperparameters in configs/mask_rcnn/mask_rcnn_r50_zino.py

total_epochs = 100
checkpoint_config = dict(interval=1)
log_config = dict(interval=50, hooks=[dict(type='TextLoggerHook')])
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
workflow = [('train', 1)]
classes = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car',
           'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike',
           'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor')
work_dir = './work_dirs/mask_rcnn_r50_caffe_fpn_mstrain-poly_1x_coco_posk_pretrain_200'
gpu_ids = range(1, 1)

Train models

To train models, run following commands.

$ python3 mmdetection/tools/train.py configs/mask_rcnn/mask_rcnn_r50_zino.py

This project used Pre-train model. But according the mmdetection doc, it used pre-train backbone restnet-50 on ImageNet.

The expected training times are:

Model	GPUs	Image size	Training Epochs	Training Time	mAP
Fast_RCNN	1x NVIDIA GTX 1080	1333, 800	100	6 hours	0.

Muti-GPU Training

I didn't test muti-GPU training.

Inference

Inference single images

you need open mmtosummit.py to change model path and output name

$ python3 mmtosummit.py

result

mmdetection Modify

Reference:

• Mask R-CNN Paper
• open-mmlab/mmdetection