ViFiT

Repository of our paper accepted in MobiCom 2023 ISACom Workshop:

Bryan Bo Cao, Abrar Alali, Hansi Liu, Nicholas Meegan, Marco Gruteser, Kristin Dana, Ashwin Ashok, Shubham Jain, ViFiT: Reconstructing Vision Trajectories from IMU and Wi-Fi Fine Time Measurements, 2023 The 29th Annual International Conference On Mobile Computing And Networking (MobiCom), 3rd ACM MobiCom Workshop on Integrated Sensing and Communication Systems for IoT (ISACom).

ISACom '23: Proceedings of the 3rd ACM MobiCom Workshop on Integrated Sensing and Communications SystemsOctober 2023 Pages 13–18 https://doi.org/10.1145/3615984.3616503

arXiv 2310.03140

Vi-Fi Dataset

New 01/16/2024: We released the synchronized version (RAN4model_dfv4p4) of our data for future usage. This version is convenient for your research without undergoing preprocessing the raw data again. Check out the details in the DATA.md file.

Official Dataset (Raw Data) link

paperswithcode link

Abstract

Tracking subjects in videos is one of the most widely used functions in camera-based IoT applications such as security surveillance, smart city traffic safety enhancement, vehicle to pedestrian communication and so on. In the computer vision domain, tracking is usually achieved by first detecting subjects, then associating detected bounding boxes across video frames. Typically, frames are transmitted to a remote site for processing, incurring high latency and network costs. To address this, we propose ViFiT, a transformer-based model that reconstructs vision bounding box trajectories from phone data (IMU and Fine Time Measurements). It leverages a transformer's ability of better modeling long-term time series data. ViFiT is evaluated on Vi-Fi Dataset, a large-scale multimodal dataset in 5 diverse real-world scenes, including indoor and outdoor environments. Results demonstrate that ViFiT outperforms the state-of-the-art approach for cross-modal reconstruction in LSTM Encoder-Decoder architecture X-Translator and achieves a high frame reduction rate as 97.76% with IMU and Wi-Fi data.

Motivation

Two types of challenges using vision-only methods: (a) Frame Drop, an entire frame in the next timestamp is not available (e.g. due to temporal down sampling to save network bandwidth, network losses, etc.), resulting in missing visual information for estimating object of interests’ detections (cyan); (b) Salient Part Missing: salient parts of objects are missing due to occlusion in the environment (purple) such as the truck or moving out of the camera’s view (orange). Missing parts are displayed in lower opacity by dotted lines. Each color represents one identity of subject of interest. Detection ground truths are shown by solid bounding boxes.

System Overview

Learning lightweight phone sensor data with rich motion information by a transformer model to reconstruct trajectories in long missing frames, which reduce the volume of data transmitted via network.

Vi-Fi Transformer (ViFiT)

ViFiT consists of multimodal Encoders for (T_c⁰, T_i and T_f ) to extract features and Vision Decoder to reconstruct the whole visual trajectory of T_c′ for the missing frames in a window with length WL. Note T_c⁰ denotes a vision tracklet with first frame only and H denotes representation dimension.

Vi-Fi Transformer (ViFiT) Architecture. ViFiT is comprised of multimodal Encoders for (T_c⁰, T_i and T_f ) depicted on the left side in parallel displayed with various degrees of opacity, as well as a Vision Decoder on the right. Information flow starts from the bottom left corner, where each tracklet for one modality (T_c⁰, T_i or T_f ) is fed into its own Encoder independently, including B blocks of transformer modules with Multi-head Self-attention (MSA). In the next step, Encoders generate multimodal representations, fused by concatenation (X_c′, X_i′, X_f′ ) and are fed into the Vision Decoder to output bounding boxes (Tc′) in missing frames.

Result

Samples of reconstructed vision tracklets T_csub>′ and ground truths GT decorated in lighter (1st and 3rd rows) and darker colors (2nd and 4th rows), respectively (Best view in color). Indoor scene is shown in the 1st column while outdoor scenes are displayed from the 2nd to the 5th columns.

Dataset

Dataset for Model - dfv4.2

Download RAN4model_dfv4.2 from Google Drive or OneDrive and follow the folder structure:

ViFiT
  |-Data
     |-checkpoints
     |-datasets
        |-RAN4model_dfv4.2
  |-src
     |-...

Pre-trained Models

Pre-trained models trained by DIoU loss can be downloaded in Google Drive or OneDrive.

Code Instructions

It is recommended to use Docker in this work. I have provided the environments in the bryanbocao/vifit container. Usage:

Pulling Docker Image

docker pull bryanbocao/vifit

Check Images

docker image ls

REPOSITORY                                                               TAG       IMAGE ID       CREATED         SIZE
bryanbocao/vifit                                                         latest    6c4d67f3d122   2 months ago    13.5GB

Run and Detach a Container

docker run -d --ipc=host --shm-size=16384m -it -v /:/share --gpus all --network=bridge bryanbocao/vifit /bin/bash

Show Running Containers

docker ps -a

CONTAINER ID   IMAGE                                                                           COMMAND       CREATED        STATUS                    PORTS     NAMES
489616f0a862   bryanbocao/vifit                                                                "/bin/bash"   5 days ago     Up 5 days                           cranky_haibt

Enter a Container

docker exec -it <CONTAINER_ID> /bin/bash

In this example:

docker exec -it 489616f0a862 /bin/bash

Train ViFiT

Under the src/model_v4.2 folder inside the container created by the commands above. Note that you need to specify <MACHINE_NAME>.

python3 Xformer_IFcC2C.py -ud -n -rm train -te 500 -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 0 -tsid_idx 5 -lw 30 -lf DIOU
python3 Xformer_IFcC2C.py -ud -n -rm train -te 500 -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 1 -tsid_idx 0 -lw 30 -lf DIOU
python3 Xformer_IFcC2C.py -ud -n -rm train -te 500 -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 2 -tsid_idx 13 -lw 30 -lf DIOU
python3 Xformer_IFcC2C.py -ud -n -rm train -te 500 -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 3 -tsid_idx 8 -lw 30 -lf DIOU
python3 Xformer_IFcC2C.py -ud -n -rm train -te 500 -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 4 -tsid_idx 4 -lw 30 -lf DIOU

Test ViFiT

python3 Xformer_IFcC2C.py -ud -n -rm test -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 0 -tsid_idx 5 -lw 30  -lf DIOU -ld_tr_eid -tr_eid 420 -ffo -mrf -w_s 29
python3 Xformer_IFcC2C.py -ud -n -rm test -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 1 -tsid_idx 0 -lw 30  -lf DIOU -ld_tr_eid -tr_eid 204 -ffo -mrf -w_s 29
python3 Xformer_IFcC2C.py -ud -n -rm test -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 2 -tsid_idx 13 -lw 30 -lf DIOU -ld_tr_eid -tr_eid 165 -ffo -mrf -w_s 29
python3 Xformer_IFcC2C.py -ud -n -rm test -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 3 -tsid_idx 8 -lw 30 -lf DIOU -ld_tr_eid -tr_eid 204 -ffo -mrf -w_s 29
python3 Xformer_IFcC2C.py -ud -n -rm test -nan linear_interp -tr_md_id Xformer_IFcC2C -m <MACHINE_NAME> -sc 4 -tsid_idx 4 -lw 30 -lf DIOU -ld_tr_eid -tr_eid 248 -ffo -mrf -w_s 29

Citation

ViFiT BibTeX:

@inproceedings{cao2023vifit,
  title={ViFiT: Reconstructing Vision Trajectories from IMU and Wi-Fi Fine Time Measurements},
  author={Cao, Bryan Bo and Alali, Abrar and Liu, Hansi and Meegan, Nicholas and Gruteser, Marco and Dana, Kristin and Ashok, Ashwin and Jain, Shubham},
  booktitle={Proceedings of the 3rd ACM MobiCom Workshop on Integrated Sensing and Communications Systems},
  pages={13--18},
  year={2023}
}

Vi-Fi (dataset) BibTex:

@inproceedings{liu2022vi,
  title={Vi-Fi: Associating Moving Subjects across Vision and Wireless Sensors},
  author={Liu, Hansi and Alali, Abrar and Ibrahim, Mohamed and Cao, Bryan Bo and Meegan, Nicholas and Li, Hongyu and Gruteser, Marco and Jain, Shubham and Dana, Kristin and Ashok, Ashwin and others},
  booktitle={2022 21st ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN)},
  pages={208--219},
  year={2022},
  organization={IEEE}
}

@misc{vifisite,
  author        = "Hansi Liu",
  title         = "Vi-Fi Dataset",
  month         = "Dec. 05,", 
  year          = "2022 [Online]",
  url           = "https://sites.google.com/winlab.rutgers.edu/vi-fidataset/home"
}

Reality-Aware Networks Project Website

Acknowledgement

This research has been supported by the National Science Foundation (NSF) under Grant Nos. CNS-2055520, CNS1901355, CNS-1901133.