mmdetection is an open source toolbox for target detection based on PyTorch. Getting started tutorial

Installation Environment

MMDetection supports running on Linux, Windows and macOS. It requires Python 3.7 and above, CUDA 9.2 and above and PyTorch 1.8 and above.

1. Install dependencies

Step 0. Download and install Miniconda fromthe official website.

Step 1. Create and activate a conda environment.

conda create --name openmmlab python=3.8 -y
conda activate openmmlab

Step 2. Install PyTorch based on PyTorch official instructions.

On GPU platforms:

conda install pytorch torchvision -c pytorch

On CPU platforms:

conda install pytorch torchvision cpuonly -c pytorch

Step 3. Use MIM Ansou MMEngine. MMCV 和 

pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"

Note:  In MMCV-v2.x, mmcv-full is renamed mmcv, if you want to install Contains a simplified version of the CUDA operator, which can be installed through mim install "mmcv-lite>=2.0.0rc1" .

Step 4. Install MMDetection.

Option a: If you develop and run mmdet directly, install it from source:

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .
# "-v" 指详细说明，或更多的输出
# "-e" 表示在可编辑模式下安装项目，因此对代码所做的任何本地修改都会生效，从而无需重新安装。

Option b: If you use mmdet as a dependency or a third-party Python package, install it using MIM:

mim install mmdet

2. Verify installation

To verify that MMDetection is installed correctly, we provide some sample code to perform model inference.

Step 1. We need to download the configuration file and model weight file.

mim download mmdet --config rtmdet_tiny_8xb32-300e_coco --dest .

The download will take a few seconds or longer, depending on your network environment. Once completed, you will find two files rtmdet_tiny_8xb32-300e_coco.py and rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth in the current folder.

 Step 2. Inference verification.

Option a: If you installed MMDetection through source code, run the following command directly to verify:

python demo/image_demo.py demo/demo.jpg rtmdet_tiny_8xb32-300e_coco.py --weights rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth --device cpu
cuda:
python demo/image_demo.py demo/demo2(1).jpg rtmdet_tiny_8xb32-300e_coco.py --weights rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth --device cuda

python demo/image_demo.py demo/demo.jpg rtmdet_tiny_8xb32-300e_coco.py --weights rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth --device cpu

 

 

 

You will see a new image  in the outputs/vis folder of the current folder. The image contains the detection frame predicted by the network. . demo.jpg

Option b: If you installed MMDetection via MIM, then open your Python parser and copy and paste the following code:

from mmdet.apis import init_detector, inference_detector

config_file = 'rtmdet_tiny_8xb32-300e_coco.py'
checkpoint_file = 'rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth'
model = init_detector(config_file, checkpoint_file, device='cpu')  # or device='cuda:0'
inference_detector(model, 'demo/demo.jpg')

You will see a list containing DetDataSample , and the prediction results are in pred_instance , including detection boxes, categories and scores.

Target detection + coverage

mim download mmdet --config mask-rcnn_r101_fpn_2x_coco --dest models

python demo/image_demo.py demo/demo.jpg configs\mask_rcnn\mask-rcnn_r101_fpn_2x_coco.py  --weights models/mask_rcnn_r101_fpn_2x_coco_bbox.pth --device cuda

3. Quick overview of mmdetection algorithm

albu_example - 示例使用Albumentations数据增强库
atss - Anchor-free一阶段目标检测模型
autoassign - 自动分配样本到计算资源的示例
boxinst - BoxInst目标实例分割模型
bytetrack - 基于ByteTrack的多目标跟踪
carafe - CARAFE通道注意力模块
cascade_rcnn - Cascade R-CNN级联RCNN目标检测模型
cascade_rpn - CascadeRPN用于Faster R-CNN的级联RPN
centernet - CenterNet中心点检测模型
centripetalnet - CentripetalNet边缘眼动检测模型
cityscapes - Cityscapes城市场景数据集
common - 通用配置和脚本
condinst - 基于条件CondiInst目标实例分割
conditional_detr - 基于DETR的条件目标检测
convnext - ConvNeXt图像分类模型
cornernet - CornerNet角点检测模型
crowddet - 群众密集场景检测模型CrowdDet
dab_detr - DAB-DETR对抗学习增强的DETR
dcn - 可变形卷积网络
dcnv2 - 可变形卷积网络v2
ddod - DDOD端到端目标检测
deepfashion - DeepFashion人体解析数据集
deepsort - DeepSORT深度学习多目标跟踪
deformable_detr - 基于可变形卷积的DETR
detectors - 通用目标检测器配置
detr - DETR (DEformable DEtection TRansformer)
dino - DINO自监督预训练模型
double_heads - Double-Heads双头目标检测
dsdl - DSdL场景文本检测
dyhead - DyHead动态头注意力
dynamic_rcnn - Dynamic R-CNN动态RCNN
efficientnet - EfficientNet图像分类网络
empirical_attention - Empirical Attention注意力机制
faster_rcnn - Faster R-CNN两阶段目标检测模型
fast_rcnn - Fast R-CNN较早的两阶段目标检测模型
fcos - FCOS全景分割目标检测
foveabox - FoveaBox凝视预测模块
fpg - Feature Pyramid Grids
free_anchor - FreeAnchor自由锚框检测
fsaf - Feature Selective Anchor-Free模块
gcnet - GCNet场景图卷积网络
gfl - Generalized Focal Loss
ghm - Gradient Harmonizing Mechanism
glip - Global Local Image Pyramid
gn - Group Normalization
gn+ws - Group Normalization + Weight Standardization
grid_rcnn - Grid R-CNN网格RCNN
groie - Gradient-weighted R-CNN Object IoU Estimation
guided_anchoring - Guided Anchoring定向锚框
hrnet - High-Resolution Network高分辨率网络
htc - Hybrid Task Cascade模块
instaboost - Instance Boostraping样本选择算法
lad - Lightweight ADetector轻量级检测器
ld - Localization Distillation知识蒸馏模块
legacy_1.x - 早期MMDetection 1.x版本配置
libra_rcnn - Libra R-CNN均衡RCNN
lvis - LVIS大词汇数据集
mask2former - Mask2Former 基于transformer的实例分割
maskformer - MaskFormer transformer based实例分割
masktrack_rcnn - MaskTrack R-CNN视频实例分割跟踪
mask_rcnn - Mask R-CNN实例分割模型
misc - 其他独立模块
ms_rcnn - Multi-Scale RCNN多尺度RCNN
nas_fcos - NAS-FCOS神经结构搜索FCOS
nas_fpn - NAS-FPN神经结构搜索特征金字塔
objects365 - Objects365数据集
ocsort - 一种基于检测的跟踪方法
openimages - OpenImages数据集
paa - Pooling-based Anchor Assignment
pafpn - Path Aggregation Network
panoptic_fpn - Panoptic FPN全景分割FPN
pascal_voc - PASCAL VOC数据集
pisa - Prime Sample Attention采样注意力
point_rend - PointRend点分割
pvt - Pyramid Vision Transformer金字塔视觉transformer
qdtrack - Quality Aware Network for Multiple Object Tracking
queryinst - QueryInst基于query的实例分割
regnet - RegNet网络结构
reid - 人员重识别模型
reppoints - RepPoints角点检测
res2net - Res2Net网络结构
resnest - ResNeSt网络结构
retinanet - RetinaNet单阶段目标检测模型
rpn - Region Proposal Network
rtmdet - Real-time Multi-scale Detector实时多尺度检测器
sabl - Side-Aware Boundary Localization
scnet - SCNet场景解析模型
scratch - 从零开始训练配置
seesaw_loss - Seesaw Loss
selfsup_pretrain - 自监督预训练模型
simple_copy_paste - Simple Copy-Paste数据增强方法
soft_teacher - Soft Teacher Semi-Supervised Object Detection
solo - Segmenting Objects by Locations单阶段实例分割
solov2 - SOLOv2
sort - SORT简单联合检测和跟踪算法
sparse_rcnn - Sparse R-CNN稀疏RCNN
ssd - SSD单阶段目标检测模型
strongsort - StrongSORT强化的SORT算法
strong_baselines - 一些强基准模型配置
swin - Swin Transformer
timm_example - 使用timm库的示例
tood - TOOD场景文本检测器
tridentnet - TridentNet三叉网络
vfnet - VarifocalNet变焦点网络
wider_face - WIDER FACE人脸数据集
yolact - YOLACT实时实例分割
yolo - YOLO系列目标检测模型
yolof - YOLOF快速Yolo模型
yolox - YOLOX优化的Yolo模型
base - 基础模块和脚本

Target Detection

Image object detection

Video detection

Command Line

mim download mmdet --config faster-rcnn_r101_fpn_2x_coco --dest models
python demo/webcam_demo.py configs/faster_rcnn/faster_rcnn_r101_fpn_2x_coco.py models/faster_rcnn_r101_fpn_2x_coco_bbox.pth --file 漫步在闵行莘庄老街道.mp4

import argparse
import os
import cv2 as cv
import torch
import argparse
import numpy as np

from mmdet.apis import inference_detector, init_detector

file_path = __file__
dir_path = os.path.dirname(file_path)
output_video_path = os.path.join(dir_path, 'result.mp4')


def main():
    args = {"file": "Nan", "checkpoint": "Nan", "config": "Nan",
            "out": "Nan", "device": "Nan", "camera_id": "Nan", "score_thr": "Nan", }
    args = argparse.Namespace(**args)
    # 自定义输入    args.device = "cuda:0"
    args.file = '/home/sha/PycharmProjects/mmdetection/workdir_hurricane/videos_test/video5.MP4'
    args.checkpoint = '/home/sha/PycharmProjects/mmdetection/workdir_hurricane/traffic_dataset_fasterRCNN/latest.pth'
    args.config = '/home/sha/PycharmProjects/mmdetection/workdir_hurricane/faster_rcnn_r101_fpn_2x_coco.py'
    args.out = 'workdir_hurricane/result.mp4'
    args.device = 'cuda:0'
    args.camera_id = 0
    args.score_thr = 0.5

    print("*" * 50)
    print(args)
    print("*" * 50)

    if not args.file:
        print('No target file!')
        exit(0)

    device = torch.device(args.device)

    print('device:', args.device)

    model = init_detector(args.config, args.checkpoint, device=device)

    camera = cv.VideoCapture(args.file)

    camera_width = int(camera.get(cv.CAP_PROP_FRAME_WIDTH))
    camera_hight = int(camera.get(cv.CAP_PROP_FRAME_HEIGHT))

    print(camera_hight, camera_width)
    fps = camera.get(cv.CAP_PROP_FPS)

    video_writer = cv.VideoWriter(args.out, cv.VideoWriter_fourcc(*'mp4v'),
                                  fps, (camera_width, camera_hight))

    count = 0

    print('Press "Esc", "q" or "Q" to exit.')
    while True:
        torch.cuda.empty_cache()
        ret_val, img = camera.read()
        if ret_val:
            if count < 0:
                count += 1
                print("Write {} in result Successfuly!".format(count))
                continue
            result = inference_detector(model, img)
            print("*" * 50)
            print(result)
            print("*" * 50)
            result_int = result[1][0:3]
            result_int = result_int.astype(int)        
            left_top = (result_int[0][0], result_int[0][1])
            right_bottom = (result_int[0][2], result_int[0][3])
            
            cv.rectangle(img, left_top, right_bottom, (0, 0, 255), 2)
            cv.imshow("img", img)
            # cv.resizeWindow("img",300,300)
            cv.waitKey(0)
            cv.destroyWindow("img")

            ch = cv.waitKey(1)
            if ch == 27 or ch == ord('q') or ch == ord('Q'):
                break

            frame = model.show_result(img, result, score_thr=args.score_thr, wait_time=1, show=False)
            cv.imshow('frame', frame)
            if len(frame) >= 1:
                video_writer.write(frame)
                count += 1
                print("Write {} in result Successfuly!".format(count))

        else:
            print('Load fail!')
            break
    camera.release()
    video_writer.release()
    cv.destroyWindow()


if __name__ == '__main__':
    main()