The YouHome ADL dataset is introduced on this page. The YouHome ADL dataset was gathered in two households that had YouHome-nodes installed. It includes 20 users' 31 common daily living activities. The dataset has a resolution of 640 x 480 and includes image, illuminance, temperature, humidity, motion, and sound sensor data. There are isolated daily activities, continuous daily activities sequences, and multiple-user interactive data. The image data were sliced from videos. The subjects' faces have been blurred for privacy reasons. At the moment, labeled images and sensor data are available.
Novel multi-user comprehensive data and annotations are under preparation and will come out soon!
- Features
- People
- Tutorial and Examples
- YouHome-nodes
- How to Cite?
- Related Repos
- Question and Comments
The YouHome ADL dataset aims to push multiple state-of-the-art smart home tasks forward. The YouHome dataset includes 20 human subjects with and without interaction performing 31 daily activities in two home environments. In all data samples, light conditions are designed to be varied. All human subjects in the dataset are labeled with user ids and bounding box coordinates. The multi-sensor data can feature many tasks, while the below are some representative examples.
Object detection is one of the most famous computer vision tasks; however, there is a lack of dataset focusing on human object detection in the home environment where the light condition can change frequently, and the occlusion of humans can happen suddenly. This differs from outdoor object detection, where the difficulty is majorly on the small size of the object or the similarity between the object and the background color. The YouHome dataset features new challenges in object detection, not only in light condition changes or occluded data but also in human objects in different poses completely different from ordinary standing postures.
Person re-identification is a follow-up task after a human object is detected. The major focus of the person Re-Id is the re-identification of pedestrians in an outdoor environment. Because most of the available datasets are surveillance videos from public places, the posture of pedestrians is mostly limited to walking. In the YouHome dataset, we present a novel concept and challenge for re-ID: the re-identification of indoor users, where users perform different postures than walking on various outfits. This adds more difficulty in recognizing users with color features. When completing the movement in the home environment, users adopt a walking posture and sitting, lying, and other irregular postures. In the previous dataset, few provided data on the re-recognition of various garments. In addition to including this content, we also included the body shape and appearance changes of the same users.
The YouHome ADL dataset possesses many unique features when performing activity recognition. First, it is a bias-free dataset that eliminates the possibility of recognizing activities based solely on their environmental context. Multiple activities are conducted and collected at each location in the YouHome dataset; thus, there is no one-to-one mapping between an activity and a context.
Existing datasets are collected through either monitoring or shooting; the YouHome dataset combines these two techniques and delivers more comprehensive data. In addition, few ADL datasets include interactive events, whereas subjects in our dataset not only follow prompts to interact during shooting sessions but also act freely without cues while we simply monitor and collect data. This creates a unique chance to train on clean data while testing on more realistic data. In both households, there are five functional areas: Entrance, Kitchen, Dining Room, Living Room, and Bedroom. Users are expected to perform different activities in each area. In addition to the visual images, ambient sensor data are also available for providing additional information.
Furthermore, the opportunity is provided to identify compound events. In the traditional multi-class classification task, only a single label can be output, whereas our multi-label compound events must output an arbitrary number of labels. In contrast to the multi-label classification of object detection, where the feature of different objects is utterly different, the multi-label classification of compound events is difficult due to the similarity of user action postures.
Cross camera-view provides a more heuristic view of the household environment and enables training and testing on more flexible settings. We provide a sufficient number of camera perspectives so that cross-view challenges can be conducted in a single household or even a single room with various combinations of view angles. Machine learning models can be trained and tested with different combinations of camera views to resemble more realistic scenarios.
This offers a chance to test the portability of the machine learning model with a different partition of training and testing set. With training data in one room and testing data in others, the performance of the model in a new home environment can be tested. In addition, by training the model in one house and testing it in the other, a more practical testing scenario can be achieved. The cross-view challenges can be performed in all tasks mentioned above; the camera_id is provided along with other labels.
Junhao Pan, Ph.D. Student at University of Illinois at Urbana-Champaign
Zehua(Neo) Yuan, Ph.D. Student at University of Illinois at Urbana-Champaign
Dr. Xiaofan Zhang, Ph.D. from University of Illinois at Urbana-Champaign (Now at Google)
Dr. Deming Chen, Abel Bliss Professor of Engineering at University of Illinois at Urbana-Champaign
In this section, we provide training and testing code for activity recognition and human re-identification. During our tutorial, the dataset is split into 8:1:1, Train: Val: Test. The activity recognition is origin from Pytorch-resnet, the re-identification is origin from Person-reID-baseline-pytorch.
- Python >3.6
- GPU Memory >= 6G
- Numpy
- Pytorch 1.10+
- Install Pytorch from http://pytorch.org/
- Install Torchvision from the source
git clone https://github.com/pytorch/vision
cd vision
python setup.py install
Download YouHome-ADL-Dataset
Preparation: For activity recognition, put the images with the same activity_id in one folder.
For human re-identification, put the images with the same user_id in one folder.
For the cross-camera challenge, further divide the dataset along with camera_id.
Train an activity recognition model by
python youhome.py --data_dir {Data_path} --save --save_folder {Save_folder} --epochs {epochs} --batch_size {batch_size} --learning_rate {learning_rate} --decay {decay} --which_gpus {which_gpus} --num_classes {num_classes} --resume {pretrained_model}
--gpu_ids which gpu to run.
--name the name of model.
--data_dir the path of the training data.
--train_all using all images to train.
--batchsize batch size.
--erasing_p random erasing probability.
Train a model with random erasing by
python train.py --gpu_ids 0 --name ft_ResNet18 --train_all --batchsize 32 --data_dir your_data_path --erasing_p 0.5Test an activity recognition model by
python youhome.py --data_dir ${data_dir} --save --save_folder {Save_folder} --batch_size ${batch_size} --which_gpus ${which_gpus} --test
Use trained model to extract feature by
python test.py --gpu_ids 0 --name ft_ResNet50 --test_dir your_data_path --batchsize 32 --which_epoch 59--gpu_ids which gpu to run.
--batchsize batch size.
--name the dir name of trained model.
--which_epoch select the i-th model.
--data_dir the path of the testing data.
python evaluate.pyIt will output Rank@1, Rank@5, Rank@10 and mAP results.
You may also try evaluate_gpu.py to conduct a faster evaluation with GPU.
For mAP calculation, you also can refer to the C++ code for Oxford Building. We use the triangle mAP calculation.
The YouHome-node is a raspberry pi embedded with a pi camera, temperature and humidity sensor, light sensor, microphone, and motion sensor. The pipelined code to use pi-node to collect data is also included in this repository. The pipeline can be run by
bash data_collection.sh {username} {eventname} {eventnumber} {timeinterval}
Please cite the following paper for the YouHome ADL dataset; it also reports the result of the baseline model.
@article{youhome2022,
title={YouHome System and Dataset: Making Your Home Know You Better},
author={Pan, Junhao and Yuan, Zehua and Zhang, Xiaofan and Chen, Deming},
journal={IEEE International Symposium on Smart Electronic Systems (IEEE - iSES)},
year={2022}
}Please submit an issue or contact Neo directly through [email protected]. We are continuously updating the labels, more use cases are coming out!!