English | 简体中文

Hospital Indoor Object Detection (HIOD) dataset

A new dataset to facilitate the task of object detection in the hospital. The proposed hospital indoor object detection (HIOD) dataset has 4,417 annotated images covering 56 object categories with 51,869 annotated object instances. The dataset is characterized by dense annotation, with an average of 11.7 objects and 6.8 object categories per image. Furthermore, a benchmark containing eight state-of-the-art object detectors is created on the dataset. The benchmark indicates that the network trained on the HIOD dataset can accurately detect and classify objects in hospitals. It is believed that the dataset and benchmark provide valuable resources for researchers and practitioners to develop computer-vision-based applications in hospitals.

For further description, please refer to this paper: Object detection in hospital facilities: A comprehensive dataset and performance evaluation

Data Preparation

The preparation of the HIOD dataset consists of four steps: object category selection, image collection, image selection, and image annotation.

Example Data

Image data were related to hospital indoor environments, such as “intensive care unit,” “operating room,” “hospital consulting room,” “hospital tour,” “hospital waiting room,” and more.

Usage

Anyone who wants to access the data folder could redirect to https://docs.google.com/forms/d/e/1FAIpQLSfI3UKkkIjvH1RGrN4BbCXCHLyRrtKt-jkJkMduw4K7ZXDNuA/viewform?usp=sf_link We Do Not Accept Any Commercial Use. So please leave your organizational email :).

The image folders contain the original jpeg files. The labels folers contains xml file of the object labels. Download both and combine them in one of the labeling softwares and then ready to be used.

The HIOD Dataset Details

Label	Number of objects	Number of images
sofa	1306	687
chair	4777	1812
foot board	606	490
overbed table	554	457
hospital bed	1581	1218
staff	1852	965
door handle	1659	1024
table	1273	796
bedside monitor	2539	1353
iv pole	1912	1166
surgical light	1138	641
breathing tube	448	406
wheel chair	60	56
patient	570	527
drawer	303	258
mouse	543	462
computer	2160	1165
bedrail	1787	813
curtain	1309	778
keyboard	730	616
infusion pump	377	281
ventilator	327	294
utility cart	877	598
panda baby warmer	94	91
visitor	196	152
dispenser	2819	1474
medical drawer	814	599
handle	5951	1382
countertop	1463	1040
cabinet	910	716
waste_bin	1233	884
faucet	717	567
TV	747	597
telephone	526	471
syringe pump	680	164
light switch	300	265
elevator panel	73	59
counter	933	737
medical waste container	218	186
push latch	330	237
operating bed	549	523
electrosurgical unit	182	158
sink	630	564
restroom assist bar	561	211
incubator	101	95
exam table	221	204
bedside table	290	249
hallway assist bar	1614	550
sequential compression	71	70
toilet	191	180
toilet handle	114	110
person	398	216
press to open	53	48
surgical instrument	87	58
xray machine	58	57
xray bed	57	51

Comparision with Other Datasets

The image shows the statistics and comparison with COCO, VOC, and OpenImages. The HIOD is found to be denser and more diverse than COCO, VOC, and OpenImages with regard to the number of objects and categories on an image. Quantitatively, our HIOD dataset has an average and median of 11.7 and 10 objects on an image, respectively. In comparison, the average number of objects per image for COCO, VOC, and OpenImages are 7.3, 8.2, and 2.7, respectively. The median number of objects per image for COCO, VOC, and OpenImages are 4, 2, and 4, respectively. On the other hand, our HIOD dataset contains an average and median of 6.8 and 6 object categories per image, respectively, which are both significantly greater than other benchmarks. The dense instances and diverse categories in the HIOD dataset lay a solid foundation to build a robust object detector in hospitals.

Benchmark Performance

Eight state-of-the-art object detectors are trained on this dataset, which is randomly partitioned into a training set (70%), a validation set (10%), and a testing set (20%). The performance of these algorithms achieved on the testing set is summarized in Table below.

Algorithm	mAP	$\mathbf{AP}_{50}$	$\mathbf{AP}_{75}$	$\mathbf{AP}_{small}$	$\mathbf{AP}_{medium}$	$\mathbf{AP}_{large}$
One-stage
YOLOv5-L	0.473	0.696	0.501	0.193	0.402	0.520
YOLOX-L	0.484	0.708	0.520	0.178	0.407	0.554
YOLOv6-L	0.517	0.737	0.553	0.211	0.418	0.597
YOLOv7	0.506	0.738	0.545	0.201	0.432	0.546
Two-stage
Faster R-CNN	0.403	0.646	0.438	0.141	0.327	0.466
Deformable DETR	0.490	0.741	0.529	0.209	0.406	0.565
VFNet	0.495	0.711	0.538	0.200	0.420	0.567
DyHead	0.430	0.645	0.468	0.145	0.345	0.507

Example Detection Results

The confusion matrices for YOLOv6-L and VFNet are shown in below figure. The diagonal value represents the recall for the corresponding object category.

Some examples of detection results by YOLOv6-L network.

Quick Start

Install

git clone https://github.com/Wangmmstar/Hospital_Scene_Data.git
pip install -r requirements.txt

Finetune on custom data

Single GPU

# P5 models
python tools/train.py --batch 32 --conf configs/yolov6s_finetune.py --data data/hiod.yaml --fuse_ab --device 0
# P6 models
python tools/train.py --batch 32 --conf configs/yolov6s6_finetune.py --data data/hiod.yaml --img 1280 --device 0

Multi GPUs (DDP mode recommended)

# P5 models
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 256 --conf configs/yolov6s_finetune.py --data data/hiod.yaml --fuse_ab --device 0,1,2,3,4,5,6,7
# P6 models
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 128 --conf configs/yolov6s6_finetune.py --data data/hiod.yaml --img 1280 --device 0,1,2,3,4,5,6,7

• fuse_ab: add anchor-based auxiliary branch and use Anchor Aided Training Mode (Not supported on P6 models currently)
• conf: select config file to specify network/optimizer/hyperparameters. We recommend to apply yolov6n/s/m/l_finetune.py when training on your custom dataset.
• data: prepare dataset and specify dataset paths in hiod.yaml
• make sure your dataset structure as follows:

├── dataset
│   ├── hiod
│      ├── images
│      │   ├── train
│      │   └── val
│      ├── labels
│      │   ├── train
│      │   ├── val

YOLOv6 supports different input resolution modes. For details, see How to Set the Input Size.

Resume training

If your training process is corrupted, you can resume training by

# single GPU training.
python tools/train.py --resume

# multi GPU training.
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --resume

Above command will automatically find the latest checkpoint in YOLOv6 directory, then resume the training process.

Your can also specify a checkpoint path to --resume parameter by

# remember to replace /path/to/your/checkpoint/path to the checkpoint path which you want to resume training.
--resume /path/to/your/checkpoint/path

This will resume from the specific checkpoint you provide.

Evaluation

Reproduce mAP on HIOD dataset with 640×640 or 1280x1280 resolution

# P5 models
python tools/eval.py --data data/hiod.yaml --batch 32 --weights yolov6s.pt --task val --reproduce_640_eval
# P6 models
python tools/eval.py --data data/hiod.yaml --batch 32 --weights yolov6s6.pt --task val --reproduce_640_eval --img 1280

• verbose: set True to print mAP of each classes.
• do_coco_metric: set True / False to enable / disable pycocotools evaluation method.
• do_pr_metric: set True / False to print or not to print the precision and recall metrics.
• config-file: specify a config file to define all the eval params, for example: yolov6n_with_eval_params.py

Inference

First, download a pretrained model from the YOLOv6 release or use your trained model to do inference.

Second, run inference with tools/infer.py

# P5 models
python tools/infer.py --weights yolov6s.pt --source img.jpg / imgdir / video.mp4
# P6 models
python tools/infer.py --weights yolov6s6.pt --img 1280 1280 --source img.jpg / imgdir / video.mp4

If you want to inference on local camera or web camera, you can run:

# P5 models
python tools/infer.py --weights yolov6s.pt --webcam --webcam-addr 0
# P6 models
python tools/infer.py --weights yolov6s6.pt --img 1280 1280 --webcam --webcam-addr 0

webcam-addr can be local camera number id or rtsp address.

• Tutorial: How to train YOLOv6 on a custom dataset

• YouTube Tutorial: How to train YOLOv6 on a custom dataset

• Demo of YOLOv6 inference on Google Colab

• Blog post: YOLOv6 Object Detection – Paper Explanation and Inference